flashcards (brainscape)

Question

What is the main trade-off when designing energy-efficient algorithms for continuous smartphone sensing? a) Accuracy vs. speed b) Accuracy vs. memory usage c) Accuracy vs. energy consumption d) Accuracy vs. internet connectivity

Answer 1

c Continuous smartphone sensing involves constantly collecting data from various sensors like accelerometers, gyroscopes, and microphones. This continuous operation drains the phone's battery significantly. Energy-efficient algorithms are designed to minimize the energy used by these sensors while still collecting usable data. Accuracy refers to how well the sensor data reflects the real world. The key trade-off lies in balancing these two aspects. Here's how: More frequent data collection (higher sampling rate) increases accuracy but consumes more energy. Less frequent data collection (lower sampling rate) conserves energy but might miss important details, reducing accuracy. Therefore, the goal is to design algorithms that can achieve an acceptable level of accuracy while minimizing the energy consumption of sensors during continuous data collection.

Answer 2

b The answer is b) Privacy concerns of users exposed to other people's sensors. Here's why: The term "second-hand smoke problem" draws an analogy to involuntary exposure. Just like inhaling smoke from someone else's cigarette, a mobile sensing system might collect data about people nearby without their consent. This scenario raises privacy concerns because sensor data can potentially reveal personal information about these bystanders.

Answer 3

c Data segmentation is for identifying and preparing the data for feature extraction in the activity recognition pipeline. In this process, we can use technics such as sliding window, or engergy based approach. Sliding window is using a window(=frame) of sample, simply slide that window with fixed overlapping and energy based approach is figure out different activities have different activity "intensities" of (or energy).

Answer 4

d Schmidt, Beigle, Gellersens has defined a context as followings : - A context describes a situation and the environment a device/user is in - A context is defined by a unique name - For each context a set of features is relevant - Context is entirely determined by the user's preferences and has no relation to the device or environment they are in. - For each relevant feature a range of values is determined(implicitly or explictly) by the context

Answer 5

D Continuous smartphone sensing, especially in the context of mobile applications, faces several technical challenges. While high computation demand and battery consumption are direct consequences of such sensing, and sensor programmability issues arise due to hardware and software diversity, user privacy through data anonymization represents a broader, systemic challenge across mobile sensing domains. It's not inherent to the continuous sensing feature but is crucial for ethical design and deployment. The focus here is on understanding the specific operational hurdles of continuous data collection and processing on smartphones, distinguishing them from overarching privacy considerations which, while vital, are managed through different mechanisms in the context of IoT data science.

Answer 6

C Persuasive technology aims to change a person's attitudes or behaviors through the use of interactive technology, while context-aware computing tailors software behavior based on the user’s current context, such as location or activity. A fitness app that tracks physical activity and encourages movement integrates both concepts by using the user’s location and past behavior (context) to motivate increased physical activity (persuasion). Unlike the other options, which may use context-awareness (A, B, D), option C specifically leverages context-aware computing to persuasively encourage a change in user behavior, aligning with the objectives of persuasive technology.

Answer 7

C Programmability focuses on managing smartphone sensors via system APIs, where controlling sensors precisely and ensuring portability across diverse operating systems and sensor models are major challenges.

Answer 8

A, B A) The app acts as a tool by providing navigation aids and monitoring safety-related behaviors (e.g., speed and route adherence), making it easier for riders to engage in safer riding practices. B) By creating a ranking system based on safety scores and rewarding points for safe riding, the app serves as a social actor, encouraging positive behavior through competition and social reinforcement. C) While the app assists in route planning and promotes safety, it does not directly provide simulated experiences or scenarios; its primary function is real-time navigation and safety feedback, not simulating different riding experiences. D) Even though the app might store data on routes and statistics, its persuasive role is not as a database but rather in its interactive features that encourage safer riding practices.

Answer 9

C Participatory sensing involves active involvement of users in collecting sensor data, as seen in the example of managing garbage cans by taking photos. Users actively participate in data collection, contributing to the complexity of operations but also influencing the quality of data.

Answer 10

C Current solutions for addressing privacy issues in mobile sensing systems include cryptography, privacy-preserving data mining, and processing data locally versus using cloud services. These solutions aim to protect user privacy by encrypting sensitive information, anonymizing data for analysis, and minimizing the transmission of personal data over the network. However, publicly sharing collected data would contradict the fundamental responsibility of respecting user privacy, as it could lead to unauthorized access or misuse of sensitive information by third parties.

Answer 11

A, D The correct options are: A) Increase the sampling rate of the accelerometer. By increasing the sampling rate, more detailed data about the user’s movements could be captured and the behavior inference could have a better accuracy. D) Use data from the gyroscope in addition to the accelerometer. Gyroscope data will help the inference process by providing additional context, such as the user’s body posture (form the device’s orientation). Additional notes: Option B could also be considered, but it is a less preferable option compared to options A and D. Relying on the users to manually label their type of activity every time they go for a walk or run is not practical and there is a good chance they will forget to label it. Option C focuses on reducing the training time, rather than directly improving accuracy in inferring the user’s type of activity

Answer 12

B, D The correct options are options B and D, as both options shows a form of automation in the data collection process, which aligns with the opportunistic sensing paradigm. Meanwhile, options A and C rely more on a manual action or input by the user, which aligns more closely to the participatory sensing paradigm

Answer 13

C The correct answer is C(Sensor Fusion). A. Proximate Selection involves considering importance to provide convenience and better understanding to users and others. B. Intervention refers to the act of becoming involved in a situation to alter, change, or influence its course or outcome. C. Sensor fusion means fusing multiple sensors to infer the user's context. D. Human-in-the-loop refers to a mode of operation in systems or processes where human involvement is integrated into the workflow. Therefore, 'C. Sensor fusion' best explains this problem.

Answer 14

C Here C, Active badge, is the context triggered action and the rest are not. a and b are Automatic Context Triggered and d is not context triggered.

Answer 15

A The right answer is a. b and d are part of "Learn" and c the part of Inform, Share, Participation

Answer 16

D The correct answer is D. To solve this problem, knowledge of participatory sensing and opportunistic sensing is required. With participatory sensing, users consciously opt to meet an application request out of their own will. Therefore, sensor data is actively collected by the user, and the quality of the data is dependent on the participants. On the other hand, with opportunisitc sensing, sensor data is automatically collected through methods such as interconnection between devices and lowers burden placed on the user. Therefore, D is incorrect in saying that the quality of data in participatry sensing is not related to the participants.

Answer 17

2 The correct answer is option 2. In the field of computer science, unique identifiers or names are frequently assigned to distinguish between different contexts within various applications. Therefore, option 2 inaccurately defines the term 'context' by suggesting it does not possess a unique name.

Answer 18

4 The correct answer to the question is option 4. 'Proximate selection' is a user interface technique that helps users choose objects or options physically close to their current location or context. In this specific scenario, selecting a nearby printer for printing tasks exemplifies 'proximate selection'.

Answer 19

a & c Persuasive technology is used to purposefully induce a change in behaviour or attitude in the user. Strategy games that have enemies that get increasingly harder to beat force the player to think more strategically, meaning that option a can be considered a persuasive technology. Using the same reasoning, option c is also a persuasive technology and option b is not a persuasive technology. For option d, since you are matched with a randomly skilled teammate, you will not be forced to think more strategically since you may get extremely high skilled players, hence it is not a persuasive technology.

Answer 20

a & d (c can be included) Participatory sensors require users to manually collect/enter data. Using this definition, option a can be trivially recognized as a participatory collection scheme, and option b can be trivially recognized as an opportunistic collection scheme. In option c, even though the user manually logs something, the collected data is the time at which they logged it as well as the temperature at that time, which is automatically collected and hence opportunistic. In option d, the user must throw their phone in order to initiate the data collection every single time, hence this is a participatory scheme.

Answer 21

B Static image analysis does not require dividing the data into smaller windows for processing.

Answer 22

C Bearing wear often leads to increased vibration. Vibration analysis can detect these changes early on, allowing for preventive maintenance before a major breakdown occurs.

Answer 23

C Cyber-physical systems - This is the defining characteristic of the Fourth Industrial Revolution, which is the focus of the question. Cyber-physical systems integrate computing, networking, and physical processes. With the advent of the internet of things (IoT), artificial intelligence (AI), and machine learning, these systems enable new ways of creating value and are a step beyond the previous revolution.

Answer 24

A & D a. This statement is false because Arduino Sense, with its 512KB of RAM, cannot support MobileNetv1, which requires 16.9MB. The statement implies that lightweight models such as MobileNetv1 are suitable for deployment on such embedded systems, which is incorrect given the hardware constraints. d. This statement is false as well. According to the lecture notes, Android’s sensing rate configuration does not guarantee the specified rates; instead, the actual sensing rate is device-dependent and varies based on operating conditions.

Answer 25

A b) this statement describes participatory sensing paradigms where citizens actively contribute data through their involvement. c)This statement suggests a specific method of data collection, focusing on external sensors rather than automatic collection from various sources. d) This statement misrepresents the purpose of opportunistic sensing, which aims to gather data from multiple sources for comprehensive analytics, rather than individualized data collection.

Answer 26

C A) False: Processing power is not the main function. B) False: Accelerometers in mobile phones have multiple uses, not limited to orienting photos. D) False: GPS sensors in mobile phones are primarily utilized for location-based services, not for adjusting screen brightness. E) False: Gyroscopes in mobile phones primarily detect device orientation for tasks like gaming and augmented reality, rather than detecting when the phone is held to the user's face during calls.

Answer 27

In time-series datasets, where the temporal order and continuity of the data points are important, interpolation is a preferred method for imputing missing values. Unlike mean or mode imputation, which might not account for the time-dependent nature of the data, interpolation uses values from neighboring data points to estimate the missing values. This method ensures that the imputed values follow the dataset's natural flow and variability over time, leading to more accurate and realistic data restoration. Winsorizing is more about limiting extreme values rather than imputing missing ones and might not be suitable for filling gaps in time-series data.

Answer 28

C Answer is (C). A) Increasing computational complexity is not a primary purpose of this technique. The sliding window method is actually a way to manage complexity by analyzing data in manageable, sequential segments. B) The technique does not inherently transform qualitative data into quantitative data, although it might be used as part of a preprocessing step that includes such transformations. C) This is correct because the sliding window technique is primarily used to analyze sequential data segments for pattern recognition, feature extraction, or smoothing purposes. D) The technique does not alter the original time series data; it's a method for analyzing the data. The original data remains intact.

Answer 29

B To solve this, you would calculate the number of readings in one window (10 readings, since the window is 10 minutes and the sampling rate is 1 reading per minute) and then consider the overlap (5 minutes, meaning 5 readings from the end of the first window are also at the beginning of the second window). Thus, the first window has 10 unique readings, and the second window also contains these 5 overlapped readings plus 5 new ones, totaling 15 unique readings in two consecutive windows when the overlap is accounted for.

Answer 30

C,D,E 20218257 A) Activity - Direct Elicitation: This is incorrect. While elicitation is used for emotions, here we're collecting activity data. Option A describes gathering emotional states, not actions. B) Emotion - Naturalistic: Not quite! Naturalistic tasks aim to evoke specific emotions, which isn't the same as observing natural emotions. Option B describes eliciting emotions, while we want to observe natural ones.

Answer 31

4 "The 3σ rule sets an upper limit at 80 (mean + 3 * standard deviation). Since 78 falls below this limit, it does not qualify as an outlier by this method." "The boxplot rule sets an upper limit at 90 (Q3 + 1.5 * IQR). Since 78 also falls below this threshold, it is not considered an outlier by this criterion either."

Answer 32

D 20218257 Chauvenet's criterion, Gaussian mixture models, and distance-based approaches detect outliers based on rarity or separation, without considering local data clustering or density. The Local Outlier Factor (LOF) assesses the degree to which a point is an outlier by comparing its local reachability distance with that of its neighbors, thus effectively detecting outliers in areas with diverse data densities.

Answer 33

C In this scenario, the EDA sensor is recording data at a consistently high frequency of 2 KHz, which aligns with the concept of polling-based fetching, where data is collected continuously at regular intervals. For ground truth labeling, the method used is Observation labeling, where experts analyze the video recordings after the fact (post-hoc) to determine the emotions of the children. This approach does not rely on real-time labeling or direct elicitation but rather on the expert analysis of observed behavior.

Answer 34

Option A is incorrect because IQR doesn't measure the spread of data around the mean. It measures the spread of data around the median. Option B is incorrect because IQR is not calculated as the difference between the minimum and maximum values. It focuses on the quartiles. Option C is incorrect because first quartile Q1 = the value under which 25% of data points are found when they are arranged in increasing order. Option D is correct becuase the IQR of a set of values is calculated as the difference between the upper and lower quartiles. First quartile Q1 = the value under which 25% of data points are found when they are arranged in increasing order. Third quartile Q3 = the value under which 75% of data points are found when arranged in increasing order.

Answer 35

C,D EDA is indeed more sensitive to psychological stimuli, making it a powerful method for emotion detection, contradicting statement a. The claim in statement b is reversed; ECG signals precede those of PPG because the heart's electrical activity happens before the blood volume changes it causes can be detected. Statements c and d are accurate, with EEG being a cornerstone in neurological and psychiatric research due to its ability to capture brain electrical activity, and EOG being valuable for measuring eye movement through electrical potential differences.

Answer 36

B Distance-based outlier detection methods, such as the k-nearest neighbors (k-NN) approach or the Mahalanobis distance, are particularly useful when dealing with data that can be modeled as a mixture of multiple normal distributions (i.e., mixture models). In such cases, outliers may deviate significantly from the expected patterns represented by the mixture components. These methods calculate the distance of each data point from the distribution centers and can effectively identify outliers within the mixture.

Answer 37

A The condition states that if a fraction of $f_{min}$ of the points are outside the distance threshold, then at least $(1−f_{min}) \times N$ points must be close (i.e., within the distance threshold).

Answer 38

B The Kalman filter predicts the current state of the system using a mathematical model of the system's dynamics. This model accounts for the previous state and any control inputs that might affect the current state. The filter estimates the uncertainty of its predictions and measurements (error covariance) and uses this estimation to weight its predictions and measurements. This process helps in effectively reducing the impact of noise in the data.

Answer 39

B ECG is widely recognized for its ability to measure the electrical activity of the heart, making it especially useful for assessing heart rate variability (HRV), among other cardiac functions. PPG is also used to measure heart rate by detecting blood volume changes in the microvascular bed of tissue, but ECG is more directly associated with heart rate variability due to its detailed capture of the electrical signals that trigger heartbeats.

Answer 40

a "Elicitation" involves collectors asking users to adhere to predefined scenarios in order to gather data. "Natural setting" entails prompting individuals to label their current activity. "Observation" refers to directly observing and recording the characteristics or situations of a subject. The Experience Sampling Method (ESM) involves sending messages to users prompting them to input their current state (label), and it is typically conducted in a natural setting. Therefore, option (a) is correct.

Answer 41

a,d The median is more robust than the mean in the presence of outliers. Outliers can significantly affect the mean, pulling it towards their extreme values. However, the median is less influenced by outliers since it depends only on the middle value of the dataset. Therefore, (a) is incorrect. An outlier is a data point that significantly differs from other data points in a dataset. Low reliability of a sensor means that precise measurements cannot be guaranteed, thereby diminishing confidence in the measurement results. Therefore, (b) and (c) are correct statements. It's not always best to remove outliers. While outliers can distort the distribution of data, they can also provide valuable insights or indicate important anomalies in the data. Removing outliers without proper justification or understanding of their origins can lead to biased or inaccurate analyses. Therefore, (d) is incorrect.

Answer 42

4 Electro-Cardio-gram (ECG): ECG measures the electrical activity of the heart. Photo-Plethysmo-gram (PPG): PPG measures blood volume changes in the microvascular bed of tissue, which can indirectly reflect changes in emotional states through variations in heart rate. However, like ECG, it is not specifically used to assess emotional arousal. Electro-Encephalo-gram (EEG): EEG records electrical activity of the brain and is crucial in neurological research and diagnosis. Electro-Dermal Activity (EDA) primarily measures the changes in the electrical conductance of the skin due to sweat gland activity. When a person experiences emotional arousal, particularly through the activation of the sympathetic branch of the autonomic nervous system, sweat gland activity increases, leading to a higher skin conductance. Therefore, EDA is used as an indicator of emotional arousal and responsiveness to psychologically significant stimuli. This physiological response is more sensitive to emotional changes rather than thermal stimuli, making it a valuable tool in assessing emotional states and reactions. Thus, the correct answer is 4.

Answer 43

c Winsorization is a method used to mitigate the impact of outliers by substituting extreme values with values closer to the rest of the data. Specifically, in a 90% winsorization, the lowest 5% of the data are replaced with the value at the 5th percentile, and the highest 5% of the data are replaced with the value at the 95th percentile. Given that the data below the 5th percentile lies between -99 and 9.75, and the data above the 95th percentile lies between 148.6 and 1001, the correct result after 90% winsorization should replace the lowest 5% of the data with 9.75 and the highest 5% with 148.6. Option 3 accurately reflects this process by substituting the appropriate values with 9.75 and 148.6, resulting in a winsorized dataset. Hence, option c) is the correct answer.

Answer 44

C Option A is incorrect because is ESM a research procedure for studying what people do, feel, and think during their daily lives. Option B is incorrect because participants need to report on their current experiences by themselves. Option C is correct because participants being prompted at random intervals to report on their current experiences Option D is incorrect because ESM involves collecting data in real-time, rather than relying on retrospective self-reports. The Experience Sampling Method (ESM): participants being prompted at random intervals to report on their current experiences involves collecting data on participants' experiences, behaviors, or thoughts in real-time as they occur in their natural environment. This method typically involves participants being prompted at random intervals to report on their current experiences, providing researchers with insights into everyday life experiences and behaviors. ESM allows researchers to capture momentary experiences and reduce recall bias that may occur with traditional retrospective self-report measures.

Answer 45

A, B Correct answers: A) and B) Option A: EDA sensors can detect changes in sweat gland activity. An increase in sweating may indicate dehydration, which could be used as an indicator for an application to suggest hydration to the user. Option B: EDA readings reflect changes in sweat gland activity, which can be correlated with stress during nightmares. With PPG, one of the metrics that can be observed is BPM, which is usually tied to emotional stress. Explanations for the incorrect options: Option C: This is typically measured with barometers. Option D: These kinds of measurements typically rely on motion sensors such as accelerometers and gyroscopes.

Answer 46

C Correct answer: C) Since the sudden drop in the measurement value is caused by mechanical failure or interference while gathering the data, this can be considered as an outlier that does not reflect the wearer's physiological state and should be removed. Explanation for the incorrect options Option A: Spikes in EDA readings can occur during intense physical activity due to increased sweat production, which is a normal physiological response and not necessarily an outlier. Option B: Lower ECG readings during sleep are expected due to the decreased heart rate as the body enters a state of rest, which can't be categorized as outliers Option D: This can be categorized as a systematic error, which can be addressed by recalibrating the device or adjusting the values based on domain knowledge

Answer 47

C Electrodermal activity (EDA), also known as galvanic skin response (GSR), is a measure of the skin's conductivity, which changes in response to sweat gland activity. This activity is primarily controlled by the sympathetic nervous system, making EDA a useful indicator of emotional arousal. When a person experiences emotional arousal, the sympathetic branch of the autonomic nervous system becomes more active, leading to increased sweat gland activity and, consequently, higher skin conductance. This physiological basis allows EDA measurements to serve as indicators of psychological or emotional states.

Answer 48

C A) Kalman Filtering: Kalman Filtering is a recursive algorithm used for estimating the state of linear dynamic systems from a series of incomplete and noisy measurements. It's not specifically designed for handling outliers in a statistical dataset. B) Chauvenet's Criterion: Chauvenet's Criterion is a rule for identifying and removing outliers from a dataset. It determines whether a data point should be considered an outlier based on the probability of its deviation from the mean, which is not the method described in the question. C) Winsorizing is a statistical transformation method used to reduce the effect of possibly spurious outliers by substituting extreme data points with less extreme ones. This could involve replacing values below the 5th percentile and above the 95th percentile with values closer to the median, thus mitigating the impact of outliers on the dataset. D) D) Local Outlier Factor (LOF): LOF is an algorithm for identifying density-based local outliers, particularly in datasets with clusters. It measures the local deviation of a data point with respect to its neighbors, aiming to identify regions of similar density.

Answer 49

C An 80% Winsorization involves replacing the lowest 10% and the highest 10% of values with the 10th and 90th percentiles respectively. In this dataset, 10% of 10 is 1, and 10% of 100 is 10. So, we replace the lowest value with the 10th percentile (20) and the highest value with the 90th percentile (50), resulting in {20, 15, 20, 25, 30, 35, 40, 45, 50, 50}.

Answer 50

B Natural observation through real-time sensors would be the most suitable method for tracking sleeping patterns using wearable devices. This method involves directly monitoring physiological signals such as heart rate, movement, and sleep stages using sensors embedded in the wearable device. It allows for continuous and accurate tracking of sleeping patterns without relying on user input or predetermined scenarios.

Answer 51

C Correct answer: C. Electrodermal Activity (EDA) is a physiological response that measures the electrical conductance of the skin, which varies with the activity of the sweat glands. Heart rate variability is often measured using electrocardiography (ECG). Electromyography (EMG) sensors are used to measure muscle tension or activity. Brain wave patterns can be monitored using Electroencephalography (EEG).

Answer 52

B Correct answer: B. The mean is calculated by summing all the values in a dataset and then dividing by the number of values. This calculation method means that every value, no matter how large or small, influences the result. When there are outliers they can significantly skew the mean. This skew can lead the mean to misrepresent the central tendency of the data, giving a distorted view of what's typical or common within the dataset. A is incorrect because outliers can significantly shift the mean away from the central mass of the data, making it a less reliable measure of central tendency in distributions with outliers. C is incorrect because the process for calculating the mean (adding all the values together and then dividing by the number of values) remains the same regardless of whether outliers are present or not. D is incorrect because outliers can either increase or decrease the mean, depending on whether the outliers are significantly higher or lower than the rest of the data.

Answer 53

C In the described case study, what you want is to setup a controlled environment with predefined scenarios for the users to perform tasks using new software. This method is known as elicitation because it actively creates situations designed to draw out specific responses or interactions from the participants. Unlike naturalistic observation, where the researcher would observe and record behaviours without intervention, or natural labelling, where users report on their activities as they change, elicitation deliberately induces a certain environment or set of circumstances to gather data on how users behave under those particular conditions. Therefore the answer is C

Answer 54

B The correct answer is B. The Kalman filter is an advanced algorithm that excels in situations where there is a need to estimate the state of a dynamic system over time. In Arnold's case, the dynamic system is his movement through space as tracked by GPS and a step counter. The Kalman filter provides a more precise and tailored approach to cleaning and imputing the data in this context, as opposed to more general methods like mean/median/mode imputation or interpolation, which do not utilize the additional information available about the system's behavior over time. Winsorizing is not as suitable since it primarily addresses extreme values rather than missing data points and does not take advantage of the dynamic model of Arnold's movements.

Answer 55

B B is true "3sigma uses the mean to calculate the standard deviation, making it more susceptible to outliers." - This statement is true. Since the mean is sensitive to outliers, and standard deviation is derived from the mean, the 3sigma method is also sensitive to outliers. This sensitivity makes it less robust compared to 3*MAD, which uses the median. A is not correct 3sigma assumes the data follows a normal distribution. If the data is not normally distributed, using 3sigma for outlier detection may not be appropriate or effective. C is incorrect because it's the opposite of the established fact that MAD is less influenced by outliers compared to the mean. MAD uses the median, which is more robust to outliers than the mean, making 3MAD more robust for outlier detection than 3sigma. D is misleading. While squaring the differences does penalize larger deviations more, it actually makes the sigma method more sensitive to outliers, not less. Outliers, which have large deviations from the mean, will have an even larger impact after being squared, making 3sigma less robust to outliers compared to 3*MAD.

Answer 56

D Correct answer is PPG. Photoplethysmography (PPG) is a simple and low-cost optical technique that can be used to detect blood volume changes in the microvascular bed of tissue. It is widely used to measure heart rate and heart rate variability (HRV) by detecting the pulse wave that travels through the blood vessels each time the heart beats. PPG sensors are commonly found in many consumer-grade wearable devices, such as fitness trackers and smartwatches, because of their cost-effectiveness and ease of integration.

Answer 57

C A) Notification schedule: This refers to the timing and frequency of notifications sent to participants. The description mentions that notifications are sent randomly six times a day, except at night, which indicates that the schedule is already a consideration B) Notification expiry: This is the window of time within which participants must respond to a notification, set at 3 minutes in the study. C) Inter-notification time: This refers to the time between consecutive notifications. Although notifications are said to be random, ensuring a minimum or maximum time between them (to avoid clustering or long gaps) is crucial for balancing data collection throughout the day and reducing participant burden. D) Study duration: The overall length of the study is mentioned as 5 days, ending next Friday. However, the practitioner didn’t consider inter-notification time, i.e. time gap between random notifications. Hence, the answer is C

Answer 58

C A) 3 sigma rule is from the fact that if X follows normal distribution with mean mu and sigma, X will fall into 3 sigma interval, mu - 3*sigma

Answer 59

D A) MAD (Median Absolute Deviation) is calculated based on the median, not the mean. MAD measures the dispersion of a dataset by calculating the median of the absolute deviations from the median. B) This statement is partially correct. MAD is indeed more robust against deviations from normality and can work well with non-normal distributions. However, standard deviation (σ) does not necessarily assume a normal distribution; it is commonly used with various distribution types. C) While MAD does provide consistent estimates of dispersion even in the presence of outliers, this statement oversimplifies the comparison with σ. Standard deviation (σ) can also be robust to outliers under certain conditions, especially if the data follows a normal distribution or if robust methods like Winsorized standard deviation are used. D) MAD focuses on the median, which is less influenced by extreme values, resulting in a more robust estimate of dispersion compared to σ, which can be inflated by outliers.

Answer 60

A, B, E A) ECG (Electrocardiogram) directly measures heart activity by detecting electrical signals produced by the heart muscle. It provides precise information about the heart's electrical conduction system and is commonly used in medical settings. B) PPG (Photoplethysmography) relies on optical measurements, capturing blood flow changes via a small LED light. PPG sensors typically measure changes in light absorption caused by blood volume changes, providing an indirect measurement of heart rate and blood flow. C) While ECG is often considered the standard for measuring heart rate due to its accuracy and direct measurement of cardiac electrical activity, PPG can also be reliable when used properly, especially in consumer-grade devices. D) PPG sensors can provide real-time heart rate measurements, and while they can calculate moving averages, they are not limited to this type of measurement. E) As A) and B) are correct, this option is also correct.

Answer 61

C The Kalman filter is a sophisticated method that is used for both detecting outliers and imputing missing values, leveraging prior knowledge about the data's process and measurement models.

Answer 62

C The Least Squares Method is primarily used for modeling relationships in regression analysis, not for detecting outliers. Chauvenet's criterion, Gaussian Mixture Models, and the Local Outlier Factor are techniques that can be used for outlier detection.

Answer 63

D In order to collect sufficient data, you need both positive and negative samples (people with good and bad form); thus A is not valid. Since human movement can vary based on a variety of factors, someone with good form might execute a movement incorrectly because they don't feel well, thus making option B invalid. Using similar logic option C is also invalid; participants might think they have the correct form without knowing that they are doing an exercise incorrectly. Option D is correct since you can see if someone performed a movement correctly, making it the most reliable.

Answer 64

D When given sinusoidal data, the mean and median imputation will produce a straight line across the middle of the sinusoidal data; although this solves the issue of missing data it doesn't do it very well. A similar problem occurs with mode imputation,except this time the straight line can occur at any point. Interpolation is the best option since it will preserve the sinusoidal shape.

Answer 65

2, 5 Pseudo sensors of smartphones can indicate the type of activity based on software. For ground truth labeling, a method of showing a video of a specific emotion and then asking the user to rate the current emotion can be used. Therefore, showing sad videos to obtain ground truth labeling is not avoided. Users may be required to follow predetermined scenarios for ground truth labeling of activities. According to the Experience Sampling Method, users can be asked to randomly label a current emotion state. It cannot be said which is always more accurate: self-labeling or labeling from observation. Therefore, 2 and 5 are incorrect.

Answer 66

2 To solve this problem, we need to know about distribution-based outlier detection. After applying the 3σ rule, the number of inliers is 10 because all values in the data lie within the range of inliers. On the other hand, after applying the 3MAD(Median Absolute Deviation) rule, '1028' is outside the range of inliers, so the number of inliers is 9. Based on this, 3MAD is more robust than 3σ.

Answer 67

b a) Incorrect. This describes the "Elicitation" method where users follow predetermined scenarios, not the "Natural" method which involves in-situ labeling by the users themselves. b) Correct. The "Natural" method involves in-situ labeling, which means asking people to label their current activity or emotion whenever there's a change of activity. c) Incorrect. This option describes the "Observation" method where labeling is done by an observer or through video recording with post-hoc labeling, not by the users themselves in their natural setting. d) Incorrect. The use of automated algorithms to label sensor data would not involve direct user input and hence does not align with the "Natural" method of ground truth labeling, which relies on user-generated labels.

Answer 68

c a) Incorrect. Winsorizing modifies extreme data points to reduce the impact of outliers but does not incorporate prior knowledge about the data, which is the key aspect of the Kalman filter's approach. b) Incorrect. A single distribution assumption for attribute noise reduction is a characteristic of distribution-based outlier detection methods, not the Kalman filter which utilizes a model-based approach. c) Correct. The Kalman filter uses models of the system's process and measurement to predict and correct state estimates, which is particularly useful in sensor data calibration and addressing outliers and missing values. d) Incorrect. Automating outlier detection based on local density and distance factors refers to other methods like the local outlier factor (LOF), which are different from the Kalman filter's model-based prediction and correction approach.

Answer 69

D ESM typically employs random or interval notification schedules, triggering prompts at various times throughout the day to capture momentary experiences. These prompts can also be event-based, such as when specific actions occur (e.g., incoming calls, app use). Option D describes a daily survey sent at a fixed time each day, which does not align with the principles of ESM. In ESM, the timing of prompts is variable and often unpredictable, aiming to capture experiences as they naturally occur rather than at pre-scheduled intervals. Therefore, a daily survey sent at a fixed time each day does not fit the criteria of ESM and is not considered an example of this method.

Answer 70

D Imputation methods are techniques used to handle missing data by replacing them with estimated values. Options A, B, and C describe typical scenarios of imputation. Option D is incorrect because it involves removing valuable data points rather than imputing missing values. This approach can lead to biased results and reduced sample sizes, potentially compromising the validity of the analysis. Therefore, discarding observations with missing values is not considered a proper imputation method.

Answer 71

C Ground truth labelling involves assigning known labels to data collected from sensors, providing a reference against which the data can be analyzed and the performance of sensing systems can be evaluated.

Answer 72

C Winsorizing is a method of limiting extreme values in the dataset to reduce the effect of potentially spurious outliers. By capping extreme values to a certain percentile at both ends of the data range (e.g., the 5th and 95th percentiles), Winsorizing reduces the influence of outliers on the analysis, leading to more robust statistical estimates.

Answer 73

B EDA measures the skin's electrical conductivity, which changes momentarily in response to various stimuli. These changes are primarily due to the activity of the sweat glands, controlled by the sympathetic nervous system, making EDA a valuable indicator of emotional arousal. This makes it a crucial measure in studies related to stress, excitement, or emotional states.

Answer 74

C Chauvenet's criterion involves determining a probability band centered on the mean of a normal distribution that should reasonably contain all samples in the dataset. It helps identify outliers by excluding data points that fall outside this band, assuming a normal distribution.

Answer 75

C Explanation: Physiological signal measurements can be affected by movement interference, making it challenging to obtain accurate data in real-world settings where participants may be moving or active.

Answer 76

B Explanation: While a small k value can be more sensitive to the local data points and potentially improve accuracy, it also makes the model more susceptible to overfitting the training data. This means the model performs well on the training data but may not generalize well to unseen data.

Answer 77

2,3 Option 2 mislabels an observation method as the elicit method for emotion data, while Option 3 inaccurately assigns an elicit method for emotion data as an observation method for ground truth labeling.

Answer 78

4 Chauvenet’s criterion is a Distribution-based outlier detection method

Answer 79

a PPG and ECG both measure heart rate. PPG measures the blood volume changes in the skin, while ECG records the electrical activity of the heart. Ultimately, both technologies are utilized to measure heart rate. Therefore, (a) is the correct answer.

Answer 80

c Chauvenet's criterion employs statistical methods for outlier detection. It assesses the extent to which data deviate from the range of standard deviation and can be applied to data following a normal distribution. However, Chauvenet's criterion does not involve excluding the largest or smallest values from a data set. Instead, it evaluates the degree to which data points deviate from a criterion to identify and potentially remove outliers. Therefore, (c) is the incorrect answer.

Answer 81

c Electrodermal Activity (EDA), also known as Galvanic Skin Response (GSR), measures the electrical conductance of the skin, which varies with its moisture level. This method is particularly useful in psychological research because the sweat glands are controlled by the sympathetic nervous system, and thus, changes in skin conductance can be indicators of emotional arousal or stress. Unlike cognitive thoughts (A) or measuring environmental factors (B), EDA provides insights into the autonomic physiological responses to emotional stimuli, which are not directly observable. This makes it valuable for studying the subconscious aspects of human emotion and stress responses, far beyond what is possible with measures of physical activity (D)

Answer 82

d In time-series datasets, where the temporal order and continuity of the data points are important, interpolation is a preferred method for imputing missing values. Unlike mean or mode imputation, which might not account for the time-dependent nature of the data, interpolation uses values from neighboring data points to estimate the missing values. This method ensures that the imputed values follow the dataset's natural flow and variability over time, leading to more accurate and realistic data restoration. Winsorizing is more about limiting extreme values rather than imputing missing ones and might not be suitable for filling gaps in time-series data.

Answer 83

B Overlapped windowing involves choosing how much windows should overlap, typically to ensure sufficient data coverage and to capture relevant information across windows. However, it also cautions that overly high overlap can lead to features being too similar, potentially causing overfitting because of the limited variation among the generated features. This concept is discussed in the context of handling numerical data in the time domain, where overlapping windows are utilized to create features from sensory data. So the answer is B

Answer 84

A Length is a numerical data type because it quantifies an amount. It's continuous as it can represent any value within a range, not just integers. When analyzing such data with Overlapped Windowing, choosing too large a window size can lead to overfitting. Overfitting occurs since features become too similar (limited variation)

Answer 85

C First, let's calculate the number of windows without overlap: (Total time window) / (Window size) = 12s / 4s = 3 windows. Now, since there's a 50% overlap, each window will overlap with the next one by 2s (half of the window size). So, in total, we'll have 5 windows: the 3 distinct windows plus two overlapping windows.

Answer 86

A Temperatures are numerical as they are continuous measurements. Smartphone brands are categorical as they represent different groups. Movie ratings are ordinal as they convey an order of preference or quality. Ages are numerical as they are discrete measurements.

Answer 87

D A and B are real life situations can use time frequency signal to extract features. C is correct. D is wrong cause F(k) = fs*k/N,k stands for how many periods in the samples.

Answer 88

3 Option 1, 2 are the advantages of distinct time window over overlapped time window. Managing the overlapping boundaries and coordinating the processing of overlapping windows can add complexity to the implementation compared to distinct time window so option 3 is incorrect. In an overlapped time window approach, consecutive windows typically overlap with each other, meaning that each data point is included in multiple windows. This overlapping allows for more frequent updates and a finer-grained analysis of the data. As a result, the temporal resolution, or the ability to capture changes in the data over time, is enhanced. Therefore, option 3 is correct

Answer 89

D Raw data often contains noise and outliers. The primary goal of feature engineering, including feature extraction, is to simplify the data representation while preserving relevant information. Features can include numeric values, strings, or categorical values. Feature extraction is essential when dealing with sensory data. Hence, the correct answer is D.

Answer 90

D a. Discrete data consists of integer values and is measured in a distinct form. The number of children is an integer value and corresponds to discrete data. b. Continuous data takes on values over an infinite range and is measured in decimal form. Weight corresponds to continuous data. c. Nominal data represents categories without any order or ranking. Hair color corresponds to nominal data. d. Ordinal data refers to data with categories that have a specific order or ranking. However, gender is treated as nominal data. Nominal data represents categories without any order or ranking

Answer 91

A The correct answer is A - Age is a continuous variable because it can take on any value within a range. - The number of cups of coffee per day is discrete because it counts the number of cups, which are whole numbers. - The favorite type of coffee is nominal data because it categorizes the coffee types without any order. - The preference ranking for coffee is ordinal data because it expresses an order of preference without specifying the exact differences between ranks.

Answer 92

A The choice of overlap ratio is crucial as it influences the level of detail captured in features and the computational resources required. A higher overlap can capture more nuanced patterns, beneficial for the model's performance but at the expense of increased computational demands.

Answer 93

D The window size parameter in feature engineering, denoted by λ, determines the length of data segments over which features are summarized, capturing essential temporal characteristics within those intervals (D). It does not directly measure noise (B). It does not solely control window overlap—overlap is related to overlap ratio rather than size (C). Contrary to statement (A), an overly large window size can actually lead to features being too similar (limited variation) and may result in overfitting, as it might not capture the necessary detail needed for the model to generalize well to new, unseen data.

Answer 94

C windows allow us to capture more nuanced patterns in the data by allowing for partial overlap between consecutive windows, which can enhance the model's ability to detect subtle features in sensory data and ensure that no data at the edges of a data segment goes missing

Answer 95

C Extracting features such as mean, standard deviation, and specifying the window size from time-domain data is crucial for understanding the underlying patterns and variability in sensory data. These features help summarize the data's central tendency and variability over specified intervals, providing meaningful insights that can significantly enhance the accuracy and efficiency of predictions made by machine learning models. This process is essential for transforming raw sensory data into a format that can be effectively utilized for analysis and decision-making

Answer 96

a) Windowing is a technique used in spectral analysis to reduce spectral leakage and improve frequency resolution. It involves dividing a signal into segments and applying a window function to each segment before performing the Fourier transform. This process helps minimize the distortion caused by spectral leakage, resulting in more accurate frequency representation in the frequency domain analysis.

Answer 97

Utilizing a smaller window size during the feature extraction process in the frequency domain tends to cause more leakage in the frequency spectrum. Leakage refers to the spreading of signal content to many frequencies rather than being confined to specific ones, which can complicate the interpretation and analysis of the frequency data. This effect is mentioned as a consequence of using smaller window sizes, as it can distort the true frequency content of the signal being analyzed.

Answer 98

4 The Fourier Transform reveals frequency components of a signal, including repetitive patterns. Option 4 contradicts the fundamental principle of Fourier analysis, which identifies repetitive patterns as distinct frequencies in the frequency domain. Therefore, option 4 is incorrect concerning data features and Fourier Transform properties.

Answer 99

C (A) points out that excessive overlap increases the overlap between data points, which can create a tendency to over-adapt to certain patterns or noise during analysis or modeling. (B) suggests that in situations where data is limited, a high overlap ratio can maximize data utilization and increase the accuracy of your analysis. (C) incorrectly suggests that using overlapping windows can prevent data duplication. Overlapping windows maintain the continuity of information and enable sophisticated analysis through overlapping data points. (D) emphasizes that overlapping windows can better preserve the boundaries of the data compared to distinct windows.

Answer 100

B Option B is correct because high overlap leads to redundant data, making features too similar and reducing variation, which leads to overfitting. Explanations for the wrong choices: Option A is incorrect because the windowing process aims to make the signals more continuous. Excessive overlap will enhance continuity, not reduce it. Option C is incorrect because spectral leakage is related to the window size and the type of window function used, not the overlapping factor. Option D is incorrect because increasing the overlap factor generally does not make it harder to locate transient events. In fact, a higher overlap can improve the detection of such events by ensuring they are captured in multiple windows.

Answer 101

B Correct Answer: B) Counting the number of students in a class represents discrete data while measuring their average height represents continuous data. Option A: Temperature readings are continuous data because the temperature can vary in a continuous range and can be measured precisely. Colour preference, while it is nominal data because it categorizes colours without a natural order, cannot be paired with temperature readings as they are different types of data. Option C: Gender is an example of nominal data, not ordinal data. Nominal data includes categories without a natural order or ranking, and gender categories typically do not have a rank order associated with them. Option D: The frequency of an activity is not continuous data but rather ordinal data if the frequency is expressed in categories such as "Never, Rarely, Sometimes, Usually, Always." If the frequency were measured in exact counts or measurements (e.g., number of times exercised per week), it could be considered discrete data. However it does not represent continuous data since frequency categories do not have an infinite range of values.

Answer 102

5 To solve this quiz, we need to understand the types of data. Temperature is continuous data because we can measure it. Income level is ordinal data since we can order it from low to high (Low < Middle < Upper). Gender is nominal data because it represents categories without inherent order or rank. The number of siblings is discrete data because we can count it as distinct, separate values. Therefore, the correct answer is option 5.

Answer 103

C Numeical data which is equivalent to Quantitative data includes Continuous data (which you can measure continuously) and Discrete data (which you can count). Also, Categorical data which is equivalent to Qualitative data includes Nomial data (which you can brand) and Ordinal data (which you can rank). Hence, the correct answer is C.

Answer 104

B Overfitting. Because too much overlapping means same data is being used over and over again, which is bad. This is also mentioned in slide 11

Answer 105

A Explanation: Dummy coding, or OneHotEncoding, is a technique used to convert nominal (categorical) variables into a binary matrix representation. This process involves creating a new binary (dummy) variable for each level of the categorical variable. In the context of machine learning, this is particularly useful because it allows models to treat each category as a separate entity or feature. This transformation is crucial for models that can only interpret numerical input, such as linear regression, where the presence or absence of a category is indicated by 1 or 0, respectively. Thus, dummy coding enables the inclusion of nominal values as distinct features in the model, facilitating the analysis of data that includes categorical variables.

Answer 106

Getting the highest amplitude frequency involves identifying the frequency component with the maximum amplitude in the frequency domain, which is a common approach in signal processing to find the dominant frequency. Frequency-weighted signal average is a method where each frequency component is weighted by its amplitude before calculating the average, providing a feature that gives more importance to frequencies with higher amplitudes. Power spectrum entropy involves calculating the entropy of the power spectrum, which is a measure of the distribution of power among different frequency components and can provide insights into the complexity or predictability of the signal. However, 4) Counting the number of data points in the frequency domain is not a meaningful feature extraction method from the frequency domain. The number of data points is simply a function of the length of the time-domain signal and the parameters of the Fourier transform, and it does not provide insightful information about the characteristics of the signal in the frequency domain.

Answer 107

b a) Incorrect. Increasing the computational complexity is not a goal of feature engineering; it aims to simplify the model's understanding of data. b) Correct. c) Incorrect. Feature engineering's purpose is not to reduce model accuracy; rather, it enhances model performance by providing relevant features. d) Incorrect. The main objective of feature engineering is not to improve data's visual appeal but to make it more actionable for predictive modeling.

Answer 108

A,B A larger window is used to drastically reduce number of data point as well as when values don't change for a long time; hence A and B are valid answers. A smaller window is desired when there is a lot of variation in a short time and when don't wnat as smooth of a curve; hence C and D are invalid answers.

Answer 109

C The purpose of applying a Hamming window (or other window functions like Hanning or Blackman) in frequency domain analysis is to smooth the abrupt ends of a signal. This technique helps in reducing spectral leakage, which occurs when energy from the main frequency component spreads to other frequencies, potentially distorting the signal's frequency spectrum

Answer 110

D Answer: D) DBSCAN does not consider temporal characteristics, resulting in clusters resembling long traces instead of distinct visiting/staying points. Explanation: The limitation described in the context is that DBSCAN ignores temporal characteristics, leading to clusters that resemble long traces rather than representing distinct visiting or staying points. This can affect the accuracy of identifying semantically meaningful places in mobility data processing.

Answer 111

B Answer: B) Re-sampling the label state data at a small time interval and utilizing forward-filling to approximate the duration. Explanation: The suggested approach involves re-sampling the label state data at a small time interval and using forward-filling to approximate the duration, which simplifies the calculation process. This method provides an approximate duration by multiplying the number of items by the unit time, offering a less cumbersome alternative to considering all cases of temporal overlap.

Answer 112

D Explanation: Density-based clustering algorithms, like DBSCAN, define clusters based on areas of high density separated by areas of low density. While they are advantageous for their ability to find clusters of arbitrary shapes and for their robustness to outliers, they can struggle in datasets where clusters have significantly varying densities. This is because the algorithm uses a single set of parameters (epsilon and minPts) to determine the neighborhood size and the minimum number of points required to form a dense region, respectively. These parameters are global and may not fit well with all clusters if their density varies widely across the dataset, leading to potential issues in accurately identifying all clusters.

Answer 113

Numeric values can be readily used for machine learning, but “nominal” values cannot be used directly. Dummy variable is a variable that can assume either one of two values (usually 1 and 0), where 1 represents the existence of a certain condition and 0 indicates that the condition does not hold. By this methods, nominal values can be used in machine learning algorithms. Therefore, the correct answer is 3

Answer 114

Options 1 and 2. are incorrect because reducing the amount of training data or increasing the model complexity typically exacerbates overfitting rather than reducing it. Option 3 is incorrect because decreasing the learning rate may slow down the training process but does not directly address overfitting. Regularization try to eliminate factors that do not impact the prediction outcomes by grading features based on importance. For example, mathematical calculations apply a penalty value to features with minimal impact. Therefore, option 3 is correct answer.

Answer 115

C The Haversine Distance is the most appropriate for measuring distances between points on the Earth's surface, represented by GPS coordinates (latitude and longitude). It accounts for the spherical shape of the Earth and provides the shortest distance between two points on the globe. Euclidean and Manhattan distances are more suited for flat Cartesian coordinates, and Cosine Similarity is typically used for measuring similarity in orientations or in text data, not for physical distances on a sphere.

Answer 116

C For time-series data, it's essential to respect the sequential nature and temporal dependencies within the data. Therefore, the dataset should be divided into k consecutive folds following the time order. Importantly, the folds selected for validation should have later timestamps than those used for training, ensuring that the validation process mimics the real-world scenario where predictions are made for future events, based on past data.

Answer 117

B The Fourier Transform (FT) is a powerful tool in frequency domain feature engineering, as it decomposes time series data into its constituent sinusoidal components. This decomposition allows for the identification of repeating patterns within the data, such as the detection of walking patterns in accelerometer data. By analyzing the amplitude and phase of these sinusoidal components (using Fast Fourier Transform, FFT), it is possible to extract meaningful features that describe the underlying phenomena or activities captured by the sensory data.

Answer 118

B Addressing imbalanced datasets effectively necessitates a holistic strategy encompassing data preparation, algorithm optimization, and appropriate evaluation methods. This includes rebalancing the dataset through techniques like oversampling or undersampling, incorporating algorithmic modifications such as cost-sensitive learning to weigh classes differently, and employing evaluation metrics like AUC-ROC or precision-recall curves that provide a nuanced understanding of model performance across different classes. Such a comprehensive approach ensures that the model is robust, fair, and performs well across both majority and minority classes.

Answer 119

A a. When converting nominal data with k levels into dummy variables, one level is used as the reference category, allowing for k - 1 degrees of freedom. b. Numerical values can also be categorized. For example, we can categorize a temperature as low, medium, and high. c. Clustering of GPS coordinate data involves identifying clusters using various methods based on defined distance metrics, which can be used for feature extraction. d. The DBSCAN algorithm is a density-based clustering technique that considers epsilon and min_points.

Answer 120

C a. Stopping training before the model learns too much from the training data is an effective method to prevent overfitting. b. Pruning in decision trees is a technique used to reduce the complexity of the model by reducing the depth of the tree or removing unnecessary branches, thereby improving the model's generalization performance. c. Using the same data repeatedly for training can lead to overfitting. In such cases, the model may perform well on the training data but struggle to generalize to new data. d. Regularization is a technique to simplify the model by selecting only useful features, thereby preventing overfitting.

Answer 121

A The correct answer is A. supervised - unsupervised - regularization. This is because feature selection often involves supervised methods where the relevance of features is determined based on their relationship with the output variable. Dimension reduction techniques like PCA (Principal Component Analysis) and auto-encoders are typically unsupervised, as they reduce data dimensions without relying on output labels. Finally, techniques like dropouts and early stopping are regularization methods used to prevent neural networks from overfitting to the training data.

Answer 122

E The correct answer is E: K-fold cross-validation. This technique involves dividing the data randomly into k segments. In each round of cross-validation, one segment is reserved for testing the model, while the remaining k-1 segments are used for training. This process is repeated k times, with each segment serving as the test set once. This method is highly effective for assessing the model's performance and its ability to generalize to unseen data. It ensures that every data point is used for both training and testing, providing a comprehensive evaluation of the model's performance.

Answer 123

C Dummy coding removes the extra degree of freedom by using only k–1 features in the representation. This is in contrast to one-hot encoding, which allows for k degrees of freedom, where k is the number of categories.

Answer 124

B The correct approach is to perform feature selection within each cross-validation fold, using only the training data for that fold and excluding the test data. This ensures that the test data remains unseen during the feature selection step, correctly mimicking the application of the classifier to an independent test set. Choice A involves performing feature selection on the entire dataset, including the test data, which violates the principle of cross-validation. Choices C and D are incorrect because feature selection should not be performed on the test data or using data from other folds.

Answer 125

C The lecture highlights the importance of identifying semantically meaningful places from GPS data for feature generation in mobility data processing. This is achieved by clustering GPS coordinates, which helps in discovering frequently visited or stayed locations. The use of a clustering algorithm, such as DBSCAN, is specifically mentioned as an approach to handle this task. Clustering allows for the transformation of raw GPS coordinates into meaningful categorical data, such as the identification of places like "home" or "work," which can significantly enhance the interpretability and effectiveness of machine learning models. This approach goes beyond simply using raw coordinates or attempting to directly calculate features such as speed or distance traveled, instead providing a layer of semantic understanding to the mobility patterns of individuals.

Answer 126

B Stratified K-fold cross-validation is chosen for datasets with uneven class distributions because it ensures each fold has the same class ratio as the entire dataset, preserving class balance across all folds. This approach provides a more accurate performance assessment by avoiding bias that might occur if some folds had significantly different class proportions than others.

Answer 127

d) Dummy coding uses k - 1 variables and usually treats the first category as the baseline (the intercept in a linear model). Therefore, the mapping is as follows: On Table = (0, 0, 0) -> intercept Washing Hands = (1, 0, 0) Standing = (0, 1, 0) Driving = (0, 0, 1)

Answer 128

a) The correct answer is a) To ensure that each fold has approximately the same proportion of different categories or classes. Stratified k-fold cross-validation is specifically designed to address the issue of class imbalance in the dataset. By ensuring that each fold maintains a similar distribution of classes as the original dataset, this method provides more accurate and representative estimates of model performance, particularly in scenarios where certain classes are underrepresented. This approach is crucial for ensuring that the model is trained and evaluated on a balanced representation of the data, leading to more reliable generalization to unseen samples. Options b), c), and d) are incorrect because they do not capture the primary purpose of stratified k-fold cross-validation, which is to address class imbalances and ensure fair evaluation of the model across different classes or categories.

Answer 129

a), c) Dummy coding is a method utilized in regression analysis to transform categorical variables into numerical form. It involves creating binary variables, known as dummy variables, for each category within the categorical variable. These dummy variables retain the categorical information while allowing for efficient inclusion in regression models. Thus, dummy coding enables the representation of categorical variables in regression analysis, facilitating the examination of their effects on the outcome variable.

Answer 130

a), c) Overfitting occurs when a model fits the training data too closely, capturing noise instead of the underlying patterns. This leads to excellent performance on the training data but poor generalization to new data. Essentially, the model memorizes the training data instead of learning meaningful relationships, resulting in inaccurate predictions on unseen data. To prevent overfitting, techniques like regularization and cross-validation are used to ensure the model generalizes well.

Answer 131

C While calculating the mean value, computing the standard deviation, and identifying the maximum and minimum values within a specified window are common methods for summarizing time-domain sensory data, determining the maximum amplitude frequency is typically a method used in the frequency domain, not the time domain. This distinction is crucial for correctly preprocessing and extracting features from sensory data, which can significantly impact the performance of machine learning models by providing them with relevant and insightful features derived from raw data.

Answer 132

B Imbalanced class distribution in a dataset can significantly bias the performance metrics of a machine learning model. For instance, in a dataset with a high imbalance ratio (where one class significantly outnumbers the other), conventional metrics such as accuracy can become misleading. A model might predict only the majority class and still achieve high accuracy, overlooking the minority class's predictive performance. This issue necessitates the use of techniques like SMOTE for rebalancing the dataset or alternative metrics such as AUC-ROC or AUC-PR that are less sensitive to class imbalance. These approaches help ensure a more accurate evaluation of the model's ability to generalize and perform across all classes, making them critical considerations in model evaluation processes.

Answer 133

C A) describes the filter method for feature selection. Filter methods use the general characteristics of the training data, independently of the learning algorithm, to evaluate the importance of features. B) describes the wrapper method for feature selection. It selects a subset of features and then iteratively tests the performance of each model. D) correctly explains the use of information gain for evaluating the importance of features. C) provides an incorrect description of Correlation-based Feature Selection (CFS). CFS considers not only the level of inter-correlation among them but also the usefulness of individual features for predicting the class.

Answer 134

D A) describes Hold-out validation, a traditional method widely used to assess a model's generalization performance. B) refers to K-fold cross-validation, where the dataset is randomly divided into k parts. For each part, the model is trained on k-1 parts and validated on the remaining one. C) points to Leave-One-Out Cross-Validation, an extreme form of K-fold cross-validation where k equals the number of instances in the dataset. D) suggests Simple Random Sampling. In this method, since the training and validation data are not separated, the validation results can be skewed.

Answer 135

C For GPS coordinates, we should use Haversine distance cause it takes into the Earth's tilt angle into consideration.

Answer 136

C The correct answer is C) Regularization methods always estimate all coefficients to be exactly zero for some predictors.

Answer 137

D The reason we use Haversine distance when measuring distance between two GPS coordiantes is because longitude and latitude give coordinate information of a point on the surface of sphere. Hence, we should use Haversine distance which measures distance between two points on the surface of sphere. If we use Euclidean distance which measures distance between points in the Euclidean space, the result can be very misleading. The statements from rest of the options are not true.

Answer 138

C One of the drawbacks of dimension reduction techniques is that it is hard to interpret the features after dimension reduction since they are linear or even nonliear combinations of original features. Hence, the answer is C. The other statements are all correct.

Answer 139

C The correct answer is C. By applying one-hot encoding, we essentially convert the categorical data into numerical data, which can then be processed similarly to typical numerical data (for instance, linear regression). Explanation for the wrong options: Option A is incorrect because interpolating categorical data is not possible. However, filling in missing data with forward or backward fill methods is still possible, albeit with careful consideration. Option B is incorrect because filtering can only be applied to numerical data. However, if we map the labels to arbitrary numerical values on a one-to-one basis, performing LPF should be possible. Option D is incorrect for a similar reason to Option B. However, proceeding with normalization after performing a one-to-one mapping is still problematic because, after scaling, the numerical data will not correspond to the initial mapping.

Answer 140

B The correct option is B. Overfitting occurs when a model is too complex, capturing noise in the training data as if it were a real pattern. Simplifying the model (by reducing parameters, pruning a few layers or neural connections, or reducing decision tree depth) will help the model generalize better to unseen data by focusing on the main trends instead of the noises. Explanation for the wrong options: Option A: Performing hold-out cross-validation is a strategy to analyze time-domain data so that the model can learn the temporal dependency of the data. It is not necessarily a way to reduce the overfitting problem. Option C: This can lead to more overfitting because by adding more features, we increase the dimensionality of the problem. This can make the model more complex and more likely to fit the noise in the training data. Option D: If a model is trained for too many epochs, it can start to learn the noise in the training data, leading to overfitting.

Answer 141

b Answer:b a) Overlapped Windowing is used in signal processing, data analysis, etc., and it involves dividing time-series data into multiple small windows. This allows for analyzing and processing various parts of the data. b) Increasing the amount of overlapping does increase the amount of data, however, too much overlapping can increase the risk of the model overfitting to the training data. c) Overlapped Windowing helps maintain the continuity of time-series data, enabling the model to recognize temporal patterns in the data and make predictions based on them. d) The size and degree of overlap of overlapping windows can indeed be adjusted based on the characteristics of the data being analyzed and the analysis objectives. Therefore, the incorrect statement is b) Increasing the ratio of overlapping can prevent overfitting by increasing the amount of data.

Answer 142

b Answer:b Overfitting occurs when a model becomes too specific to the training data and fails to generalize well to new data. Different techniques are used to tackle this problem. a) Early stopping: This method involves stopping the training process before the model overfits the training data too much, thus preventing overfitting effectively. b) Increasing network complexity: Making the model more complex often worsens overfitting rather than alleviating it. Therefore, it's not an effective method for dealing with overfitting. c) Feature selection and regularization: These techniques help reduce model complexity and prevent overfitting by focusing on important features and constraining weights. d) Expansion of training data: Providing the model with more diverse examples can help improve its ability to generalize to new data and mitigate overfitting. Hence, increasing network complexity (b) is not an effective method for dealing with overfitting.

Answer 143

D Explanation: Correlation-based Feature Selection (CFS) focuses on selecting feature subsets that are highly correlated with the class label while maintaining low inter-correlation among themselves. This approach aims to reduce redundancy and enhance the predictive power of the model. A feature that exhibits low correlation with other features avoids redundancy, and a feature that exhibits high correlation with the class label ensures its relevance and contribution to accurate model prediction. Therefore, option D is the correct choice.

Answer 144

B - Option A acknowledges the need to convert nominal data for machine learning, typically achieved through OneHotEncoding. Option C elaborates on this conversion, where each nominal category is represented uniquely. Option D underscores the direct applicability of numeric values in machine learning. - Option B is incorrect as it misstates the number of one-hot vectors needed. For a nominal feature with k labels, only k-1 vectors are necessary due to the k-1 degrees of freedom principle. For instance, with labels A, B, and C, two vectors are sufficient: [1, 0] for A, [0, 1] for B, and [0, 0] for C.

Answer 145

B Feature selection reduces the complexity of the model by eliminating irrelevant or less significant features. This enhances the model's performance by focusing on features that have more predictive power.

Answer 146

C cross-validation involves partitioning the data into subsets, training the model on some subsets (training set) and testing it on the remaining subsets (validation set), thereby providing a reliable estimate of the model's performance on unseen data.

Answer 147

2 Data-centric AI emphasizes the importance of data collection, cleaning, and preprocessing in the AI development process. While model architecture and algorithm optimization are essential components of AI development, data-centric AI recognizes that the quality and relevance of the data used to train AI models significantly impact their performance. By prioritizing data collection and preprocessing, data-centric AI aims to ensure that AI models are trained on high-quality data, leading to more accurate and reliable predictions. Therefore, correct answer is 2.

Answer 148

3 The Discrete Fourier Transform (DFT) is a mathematical technique used to transform signals from the time-domain into the frequency-domain. Audio equalization involves adjusting the amplitude of different frequency components in an audio signal. The DFT is commonly used in audio processing applications such as equalization to analyze and manipulate the frequency content of audio signals. Therefore, the correct answer is audio equalization.

Answer 149

3 Increasing samples in the Discrete Fourier Transform enhances resolution by providing more data points for better frequency distinction, crucial in fields like signal processing. Reducing samples or sampling frequency decreases resolution, while keeping samples constant doesn't improve it.

Answer 150

3 3) Mutual improvement of both model and data through iteration Explanation: Data-Centric AI emphasizes the iterative improvement of both models and data, recognizing that the quality of data directly impacts model performance and vice versa. This iterative process allows for continual refinement and optimization.

Answer 151

C "Data smells" typically refer to subtle, latent errors or issues within a dataset that are not immediately obvious. These errors are often independent of the specific context. Unlike more overt or context-dependent errors, data smells are more insidious and may only become apparent upon deeper analysis. Identifying and addressing these latent issues is crucial for ensuring the quality and reliability of data analysis.

Answer 152

B With a bandwidth of 10 Hz, the Nyquist frequency is 20 Hz. Sampling at the Nyquist frequency means the sampling rate is 20 Hz. The duration of the signal is 2 seconds, so the total number of samples is 20 × 2 = 40 samples. The frequency bins are equally spaced from 0 up to (but not including) the sampling frequency. The spacing between each bin is Sampling Frequency / Number of Samples = 20 Hz / 40 = 0.5Hz Therefore, the frequency bins will range from 0 Hz in increments of 0.5 Hz up to 19.5 Hz.

Answer 153

D Addressing data quality issues, like label errors and biases, directly impacts AI model accuracy, ensuring reliable predictions and insights. High-quality data underpins trustworthy IoT applications, making addressing these issues crucial for dependable outcomes, especially in high-stakes scenarios.

Answer 154

C Choosing a sampling rate at least twice the highest frequency of the signal, as per the Nyquist theorem, is critical to prevent aliasing. Aliasing can cause different signals to become indistinguishable when sampled, leading to frequency distortion and inaccurate signal representation. This principle ensures the integrity and reliability of signal processing in digital systems.

Answer 155

A The data cascade problem refers to the phenomenon where incorrect data can have a cascading effect on the model, leading to inaccuracies in the model's performance. Data-centric AI has emerged to address this issue, which is often ignored in model-centric approaches. Therefore, A is incorrect statement.

Answer 156

A Zero-padding in the time domain increases the length of the signal, resulting in a higher number of samples and frequency bins in the frequency domain. This increased number of bins provides a finer frequency resolution, allowing for better frequency localization. So, statement A is correct.

Answer 157

d For Data-Centric AI, we consider Human centeredness of ‘data work’ instead of AI,so d is incorrect.

Answer 158

b Data-Centric AI differs from Model-Centric AI in that it not only focuses on model improvement but also analyzes data errors, modifies data to address issues such as misfit and inconsistency, and evaluates data quality through benchmarking. In contrast, Model-Centric AI primarily concentrates on hyperparameter optimization and model refinement using the same training data without directly addressing data-related issues.

Answer 159

b Zero padding, which involves adding zeros to the existing samples before computing the Discrete Fourier Transform (DFT), enhances resolution by interpolating between existing data points. It allows us to achieve better frequency resolution without altering the original signal.

Answer 160

b Data smells are latent by definition, i.e., they represent issues that are present in the dataset and can cause problems, but are not obvious now. Data completeness should be caught at the level of data preparation, so early on in the data analysis pipeline.

Answer 161

d a) Oversampling: Oversampling involves using a higher sampling rate than necessary to capture more signal detail, which isn't the case when the sampling rate is too low. b) Nyquist theorem: The Nyquist-Shannon theorem states the minimum sampling rate required to avoid aliasing but doesn't directly explain the loss of high-frequency information due to low sampling rates. c) Frequency folding: This occurs during aliasing when higher frequencies are incorrectly represented as lower frequencies, but it doesn't directly explain the loss of high-frequency information due to low sampling rates. d) Aliasing: Aliasing happens when higher frequencies are misrepresented as lower frequencies due to low sampling rates, directly explaining the loss of high-frequency information.

Answer 162

c A modern data platform architecture for ensuring data quality typically consists of the following key layers: data ingestion, data storage and processing, data transformation and modeling, business intelligence and analytics, data observability (monitoring, tracking, triaging incidents), and data discovery and governance (documenting critical data assets). These layers collectively enable the management and maintenance of data quality throughout the data lifecycle.

Answer 163

a The FFTShift function rearranges the DFT coefficients by shifting the zero-frequency (DC) component to the center of the array. For a real-valued signal, the DFT coefficients are symmetric around the DC component. In this case, the DC component is the first element (X[0] = 2), and the remaining coefficients are arranged in alternating positive and negative frequency components. After applying FFTShift, the DC component is moved to the center, and the positive and negative frequency components are arranged in a contiguous manner on either side of the DC component. The resulting order is: [negative freq. components, DC component, positive freq. components]. Therefore, the correct answer is option a.

Answer 164

a The Nyquist sampling rate is crucial in digital signal processing to prevent aliasing. It mandates that the sampling rate must be at least twice the signal's highest frequency (Nyquist frequency) to accurately represent it digitally. Aliasing occurs with insufficient sampling rates, misinterpreting high frequencies as lower ones, causing distortion. Adhering to Nyquist's rate ensures faithful representation of analog signals without artifacts, vital for digital signal processing fidelity. Understanding and applying Nyquist's principle is essential for maintaining accuracy and fidelity in digital systems.

Answer 165

c According to Carlo Batini et al.(2009), the listed quality metrics (Consistency, Completeness, Timeliness, Validity, Uniqueness, Accuracy) represent characteristics or dimensions used to assess the quality of data. These metrics help in understanding the strengths and weaknesses of the data for different aspects such as format consistency, absence of missing values, up-to-dateness, representation of the real world, uniqueness of values, and accuracy in reflecting true values. Option C) Sampling is not a methodology for data quality assessment or improvement, nor is it a representative quality metric as outlined by Batini et al. Sampling typically refers to the process of selecting a subset of data from a larger population for analysis or evaluation. While sampling can be a technique used within some methodologies, it is not inherently a methodology itself nor a quality metric. Therefore, it is the correct choice for this question.

Answer 166

B Undersampling is a process applied to high-resolution signals to enhance computational efficiency. However, this practice risks causing aliasing, where higher frequency components can distort the signal when sampled below their Nyquist rate. To avoid this, it's essential to filter out frequencies greater than half of the undersampling frequency, Therefore we should remove signals with frequencies higher than half of the undersampling frequency.

Answer 167

D Data-centric AI emphasizes the continuous enhancement of data quality after the initial training of the model. This includes tasks such as systematic data fit improvement, ensuring data consistency, and fostering mutual growth through model-data iteration. In contrast, Model-Centric AI often focuses on refining the model through hyperparameter tuning using the same initial dataset.

Answer 168

C Data Cascades involve compounding events that cause negative downstream effects due to data issues, leading to technical debt over time. This highlights the importance of addressing data quality early in the data lifecycle.

Answer 169

C Zero padding involves adding zeros to the end of a signal to increase the number of samples, which can improve the resolution of the DFT by making the frequency bins closer together. This technique does not change the fundamental frequency of the signal but rather helps in providing a more detailed frequency analysis by interpolating between the original DFT bins.

Answer 170

D The sequence in Digital signal processing (DSP) systems is as follows: First, raw analog data is passed through an input analog filter to produce a filtered input. This filtered signal is then converted from analog to digital format using ADC. After this conversion, the digital signal undergoes digital processing. Once the processing is complete, the signal is converted back to analog format using DAC. The final step is to pass the reconstructed signal through an output analog filter for output. So the sequence is input analog filter -> ADC -> Digital Processing -> DAC -> output analog filter.

Answer 171

D The components of the Dataset Development Lifecycle include Requirements Analysis, Design, Implementation, Testing, and Maintenance. "Using" refers to the stage where the developed dataset is utilized for various tasks or analyses. While it is an essential part of the overall data lifecycle, it is not specifically a component of the dataset development process. The primary focus of the Dataset Development Lifecycle is on creating, improving, and maintaining datasets, rather than on their utilization in practical applications or analyses. Therefore, "Using" is not considered a component of the Dataset Development Lifecycle.

Answer 172

A Representative quality metrics for data consist of 6 metrics which are Consistency, Completeness, Timeliness, Validity, Uniqueness and Accuracy. Among them, "Consistentcy" is a metric of the extent to which data follows same format. Hence, the answer is A. The extent to which data correctly reflects true value is measured as "Accuracy". Rest of the options match with their original explanation. Even though not in the option, "Validity" is the extent to which data represents real-world.

Answer 173

C The aliasing refers to problem where the sampling rate is too low so that the sampled signal does not properly represent the true signal. It is recommended to sample with at least twice the original signal's bandwidth (Nyquist sampling rate). Zero padding and Resolution are concepts regarding the DFT and noise of the data is not a main cause of aliasing problem.

Answer 174

A The lecture contrasts Model-Centric AI and Data-Centric AI by highlighting that Model-Centric AI focuses on refining the algorithms or models themselves, whereas Data-Centric AI emphasizes the importance of improving the quality of the data used for training these models. This distinction underlines the shift from solely enhancing algorithmic approaches to recognizing the critical role of high-quality, well-labeled, and consistent data in achieving effective AI outcomes. Therefore, option A is the correct answer

Answer 175

A The Nyquist sampling rate should be twice the signal's bandwidth to accurately represent the signal without aliasing.

Answer 176

d Data-Centric AI aims to enhance the performance of models by prioritizing the quality and quantity of data. To improve model performance, it is essential to collect, manage, preprocess, and appropriately evaluate data. Therefore, option d) "Utilizing the same data to enhance the performance of models like Data-Centric AI" is not a description of Data-Centric AI but rather of Model-Centric AI.

Answer 177

c (a), (b): Since the bandwidth of the signal is fixed, increasing the duration of the signal does not change the Nyquist sampling rate. Thus, option (a) and (b) are incorrect. (c): Since the bandwidth of the signal is fixed, increasing the duration of the signal enhances the frequency resolution. Frequency resolution is inversely proportional to the time duration; hence, increasing the duration enhances the frequency resolution. Therefore, option c) is correct. (d): The amount of sampled data doubles as the time increases from 1 second to 2 seconds. Thus, option d) is incorrect.

Answer 178

A The resolution is given by f/N (in reality a lower value means a higher resolution), where f is the sampling rate and N is the amount of samples taken; the amount of samples itself is given by f*s, where s is sample duration in seconds; so the formula simplifies down to 1/s; this means that the only thing that will increase the resolution is increasing the sample duration, hence the answer is A.

Answer 179

C The answer is C; all of the other steps exist (and are necessary) for both Data-Centric and Model-Centric AI. In order for the model to train properly in both cases, data preprocessing must be done. In order to train the model hyperparamter and iterative optimization must be done. However, in Model-Centric AI, you only update the models, not the data.

Answer 180

B The principles of Data-Centric AI focus on systematically improving the data fit and consistency of data, emphasizing the importance of data quality over the intricacies of model design. This approach advocates for continuous and substantive interactions between AI and domain experts, aiming to enhance both the reliability of AI predictions and the overall utility of AI in real-world applications.

Answer 181

D Encryption errors are a problem for the storage of the data as well opening privacy concerns, but they do not affect the actual quality of the data.

Answer 182

B Data cascades is not a kind of measurement error, rather it is the result of accumulated errors

Answer 183

A Definition of Nyquist sampling rate.

Answer 184

c Dataset testing report is specifically designed to assess the necessity of the dataset and whether the dataset meets its requirements and is safe to use. It typically includes information about the testing procedures, test results, and any issues or concerns identified during the testing phase. By evaluating the dataset against predefined requirements and safety criteria, the testing report ensures the quality and reliability of the dataset before it is utilized in the data development life cycle. Therefore, the answer is c.

Answer 185

a, c, e To enhance frequency resolution, increasing overlap between segments (a), increasing the number of samples (c), and implementing zero-padding (e) are effective strategies. Increasing segment overlap reduces spectral leakage and effectively extends the effective window length, particularly beneficial in methods like the short-time Fourier transform (STFT). Increasing the number of samples extends the signal's duration, allowing for finer frequency distinctions in spectral analysis. Zero-padding adds zeros to the end of a signal before performing a Fourier transform, effectively increasing the data points and improving frequency resolution by interpolating between existing points in the frequency domain. Reducing the window length (b) or employing shorter time intervals (d) may improve time resolution but not frequency resolution.

Answer 186

b a) Incorrect. While iterative cycles of data cleansing are important in data management, they do not specifically refer to the compounding negative effects caused by initial data issues, which is the hallmark of Data Cascades. b) Correct. c) Incorrect. Systematically increasing the volume of data for machine learning models is a practice in data collection and augmentation, but it does not address the compounding negative effects caused by initial data problems as described by Data Cascades. d) Incorrect. Layering data from various sources can enrich the data set but does not specifically refer to the negative downstream effects that result from initial data issues, which is what Data Cascades describe.

Answer 187

a a) Correct b) Incorrect. While higher sampling rates can lead to greater resolution in the digital representation of the signal, the Nyquist rate specifically addresses the minimum requirement to avoid aliasing, rather than the resolution itself. c) Incorrect. The Nyquist rate is actually twice the maximum frequency component of the signal, not equal to it. This distinction is crucial to prevent aliasing and accurately capture the signal in its digital form. d) Incorrect. The Nyquist rate is concerned with avoiding aliasing and accurately capturing the signal, not with noise elimination. Noise reduction involves different techniques and considerations beyond just the sampling rate.

Answer 188

B To mitigate the negative impact of Data Cascades, it's crucial to implement comprehensive data quality checks and validations at each stage of the data lifecycle. This approach helps identify and address data issues early, preventing them from compounding and leading to significant technical debt and negative downstream effects on AI systems.

Answer 189

C Zero padding is a technique used in DFT to improve the frequency resolution by simulating a longer data acquisition time. It involves adding zeros to the end of the signal, which does not alter the signal's content but increases the number of points in the DFT, thereby improving the frequency resolution.

Answer 190

e To address this question, it is essential to comprehend the principles of Data-Centric AI. Data-Centric AI utilizes strategies to enhance the quality of data with each iteration. Option (a) is employed to verify errors within the data. Option (b) serves to enhance and refine the dataset. Option (c) functions as a strategy for evaluating data quality. Lastly, option (d) acts as a strategy based on data fit. Therefore, the correct answer is (e).

Answer 191

1, 4 When calculating the Nyquist sampling rate for a signal with a duration of 1 second and a fixed bandwidth of 10 kHz, it is 2*f_b, which equals 20 kHz. Also, the total number of samples is 20 kHz * 1 s = 20k, and Δf is 20 kHz / 20k = 1 Hz. Now, if the duration changes to 2 seconds, the Nyquist sampling rate remains at 20 kHz, and the total number of samples becomes 20 kHz * 2 s = 40k. Additionally, Δf becomes 20 kHz / 40k = 1/2 Hz. As the duration changes to 2 seconds, the total number of samples increases, and the resolution increases as well. On the other hand, the Nyquist sampling rate remains the same, and Δf becomes 1/2 Hz. Therefore, statements 1 and 4 are correct.

Answer 192

D (A), (B), and (C) are all valid reasons why data scientists might choose to use their own methods

Answer 193

D If only cos() is used to calculate the correlation, we can obtain only the magnitude and sign of the correlation, and we cannot obtain phase information.

Answer 194

2 Aliasing occurs when the sampling rate is too low to accurately represent the frequency content of the original signal. According to the Nyquist sampling theorem, the sampling rate must be at least twice the maximum frequency present in the signal to avoid aliasing. therefore, option 2 (A 10Hz signal sampled at 15Hz(<20HZ)) is the correct answer as it violates the Nyquist criterion and would result in aliasing.

Answer 195

3 Option 3 accurately depicts Data-centric AI by emphasizing the systematic enhancement of data consistency. Let's examine the other options. Firstly, Data-Centric AI prioritizes human-centered handling of data, making option 1 incorrect. Additionally, it entails iterative refinement of both models and data, making option 2 inaccurate. Moreover, Data-Centric AI focuses on ensuring data suitability rather than merely increasing quantity, rendering option 4 less pertinent. Therefore, option 3 remains the most fitting description of Data-centric AI.

Answer 196

B This statement accurately captures the essence of the Nyquist theorem, which states that in order to accurately represent a signal digitally, the sampling rate must be at least twice the frequency of the highest frequency component in the signal. If frequencies above the Nyquist rate are present in the signal and are not properly filtered out before sampling, aliasing occurs, leading to false lower-frequency components in the digitized signal.

Answer 197

A, C Only A correctly highlights the fundamental difference between the two AI approaches. Model-Centric AI is about refining and optimizing the models and algorithms themselves to enhance performance, which often involves tweaking model architectures or algorithm parameters specific to a task. On the other hand, Data-Centric AI shifts the focus towards the data used to train these models. It posits that by ensuring the data is of high quality and diversity, you can significantly improve the model's performance. This approach involves careful curation, cleaning, and possibly augmenting data to better train models, rather than primarily focusing on the model's internal adjustments.

Answer 198

C Data Cascades in the context of AI and data science refer to a series of compounding negative effects triggered by initial data quality issues. These cascading events can lead to a deterioration in model performance, increased costs, and delays in project timelines. Significantly, they contribute to the accumulation of technical debt, as initial shortcuts or oversights in data handling and preprocessing become more challenging to address over time. Recognizing and mitigating data cascades early in the development cycle is crucial to maintaining the integrity and reliability of AI systems.

Answer 199

A The Nyquist rate is fundamental in digital signal processing as it represents the minimum sampling rate required to accurately capture a signal without aliasing. Sampling at least at the Nyquist rate, which is twice the signal's highest frequency, ensures that the original signal can be fully reconstructed from its samples, preventing the loss of information and avoiding the misinterpretation of signal frequencies (aliasing).

Answer 200

B The Nyquist theorem states that to accurately sample a continuous signal without information loss, the sampling rate must be at least twice the highest frequency present in the signal. This prevents aliasing, ensuring that the sampled signal can be accurately reconstructed from its samples.

Answer 201

D Explanation: This statement is incorrect because Data-Centric AI actually emphasizes improving data quality (through error analysis, data cleaning, and enhancement techniques) as a primary approach to improving AI model performance, rather than focusing on adjusting the model's parameters or architecture. On the other hand, statements A and C are correct in their respective contexts, and B accurately describes a core principle of Data-Centric AI, making D the incorrect choice.

Answer 202

A Windowing in signal processing multiplies a signal with a window function, which can cause abrupt transitions at the window's start and end. This discrepancy leads to spectral leakage, where energy spreads across different frequency components rather than remaining concentrated. This spreading effect is due to the abrupt changes, causing the energy to disperse into frequencies outside the main component, altering the signal's spectrum. Thus, "start" and "end" relate to window edges, and "spreading" describes the resulting dispersion of energy.

Answer 203

d Time domain filters are generally simpler to implement and require less computational resources compared to frequency domain filters. Frequency domain filters often introduce significant processing delays due to the need for Fourier transforms and inverse transforms. Time domain filters are effective in reducing noise and unwanted signal components in real-time data streams (e.g., time domain filters can adapt their parameters in real-time based on the characteristics of the input signal).

Answer 204

D Roll-off in digital filters refers to how quickly the filter attenuates frequencies beyond the passband, specifically, the steepness or slope of the filter's attenuation in the transition band between the passband and stopband. A steep roll-off means the filter more sharply reduces the amplitude of frequencies outside its passband, which is desirable in many applications for a clear distinction between filtered and attenuated frequencies. So the clear answer is D)

Answer 205

a A low-pass filter is designed to allow frequencies below a certain cutoff frequency to pass through unaffected, while attenuating frequencies above the cutoff. This means that signals with frequencies lower than the cutoff frequency will be transmitted with little or no attenuation, while signals with frequencies higher than the cutoff will be reduced in magnitude. Therefore, the correct choice is a) Low-pass filter.

Answer 206

B Spectral leakage, an artifact that occurs when a signal is not periodic within its sample window, can distort the signal's frequency spectrum. This distortion can be mitigated by applying window functions, such as the Hamming or Blackman window, to smooth the signal's abrupt ends before performing the DFT. This process reduces the leakage by minimizing the energy spread caused by the abrupt signal cutoff, enhancing the filter's effectiveness and the accuracy of the spectral analysis.

Answer 207

D Option A - The System's Response to a Loud Pop: A loud pop is a short, sharp sound that can somewhat mimic an impulse. Option B - The System's Response to a Clap: Like a loud pop, a clap is also a brief and high-energy sound. It approximates an impulse. Option C - The System's Response to a Song: A song is a complex signal with varying frequencies, rhythms, and intensities. It bears little resemblance to an impulse, which is a singular, instantaneous event. In the context of the convolution reverb experiment, we see that a loud pop and a clap were used to characterize the building(system).

Answer 208

3 A low-pass filter is designed to attenuate or block high-frequency components of a signal while allowing low-frequency components to pass through relatively unaffected. This means that high-frequency components, such as noise or interference, are suppressed by the filter, resulting in a smoother output signal with reduced high-frequency content. Therefore, answer is 3

Answer 209

C,D Abrupt changes(Discontinuous points) in boundary could cause Spectral leakage. Since hamming window reduces the side effects of truncation by smoothing, it can reduce spectral leakage. So (a) and (b) are correct. (c) If the window size is small, truncation can occur, leading to spectral leakage. (d) Spectral leakage is not directly caused by low sampling frequency. Even at high sampling frequencies, if the signal contains frequency components that do not align well with the window boundaries, spectral leakage can still occur. Therefore, (c) and (d) are incorrect

Answer 210

C Spectral leakage occurs when a signal is truncated, leading to the spreading of its energy across other frequencies. To mitigate this effect, the slide recommends smoothing the abrupt ends of the signal by applying windowing functions such as Hamming or Blackman. These windows reduce the leakage by tapering the signal at its ends, rather than abruptly cutting it off, which is characteristic of a rectangular window.

Answer 211

B Window functions like Hamming and Hanning are commonly used in signal processing to reduce spectral leakage by smoothing abrupt ends of the signal. In the frequency domain, the input signal and filter can create an output signal through multiplication, not convolution. If the windowing size is not sufficient to contain the original signal, spectral leakage can occur. Abrupt changes or ends in the signal can result in spectral leakage. Therefore option B is incorrect.

Answer 212

C The correct answer is C. This choice is optimal because it covers the frequency range that corresponds to the typical human heart rate, which, when converted to Hertz, is 1 Hz to 3 Hz. By selecting a Band Pass Filter with a range from 0.5Hz to 4Hz, we ensure that the heart rate signals are effectively isolated from the noises happening at 10 Hz, 40 Hz, and 0.2 Hz. This range not only captures the relevant heart rate frequencies but also provides a margin to account for potential outliers.

Answer 213

B The Butterworth filter is known for its maximally flat frequency response in the passband, meaning it has no ripples. This design choice results in a gradual transition or slope as it approaches the cutoff frequency, moving towards the stopband. The slope, or roll-off, is a byproduct of attempting to maintain this flatness as closely as possible up to the cutoff frequency, beyond which attenuation begins. This gradual transition helps to reduce distortions within the passband while effectively attenuating unwanted frequencies in the stopband.

Answer 214

C To reduce passband ripple in an FIR filter, the most effective strategy is utilizing a windowing technique (C). Windowing smooths out the filter's impulse response, reducing side lobes and effectively diminishing ripple in both the passband and stopband. On the other hand, zero-padding (A) mainly enhances frequency resolution in spectral analysis without directly affecting the ripple. Decreasing the filter order (B) might simplify the filter but can increase ripple due to less precise control over the filter's frequency response. Narrowing the frequency band (D) does not directly influence passband ripple.

Answer 215

B FIR (Finite Impulse Response) filters can maintain a linear phase response, which is crucial for applications where preserving the wave shape of the original signal is important, as it ensures all frequency components of the signal are delayed by the same amount when passing through the filter. This property makes FIR filters particularly suitable for phase-sensitive applications. On the other hand, IIR (Infinite Impulse Response) filters, due to their recursive nature, can introduce phase distortion, as they may not delay all frequency components uniformly. This difference highlights the importance of choosing the right filter type based on the specific requirements of a signal processing task.

Answer 216

B Spectral leakage is a phenomenon that occurs when a signal is windowed, leading to the spreading of energy from the (mainlobe to the sidelobes. This effect reduces the spectral resolution because energy that should be concentrated in the main frequency band leaks into other frequencies. It is prompting the use of various windowing techniques to minimize the leakage and preserve as much spectral accuracy as possible.

Answer 217

B Option C) is the correct answer as it accurately interprets the frequency spectrum of a real signal, indicating the presence of fundamental and harmonic frequencies. These peaks are characteristic of periodic mechanical sources, such as the blades of a submarine's propeller, which generate distinct peaks at these frequencies due to their rotation.

Answer 218

C Increasing the order of a Butterworth filter makes its transition from passband to stopband steeper, allowing for a clearer distinction between allowed and blocked frequencies. This improves the filter's precision in separating desired signals from noise.

Answer 219

B A, C, D are all correct explanations about spectral leakage. The main reason for spectral leakage is the inappropriate window size during sampling which in turn leads to energy leaking out from the mainlobe to the sidelobes after DFT. One way to reduce this is to smooth out the abrupt ends of the windowed signal. B says that high autocorrelation leads to spectral leakage which is not the case.

Answer 220

B B is the correct answer according to the slides. For large N, the function behaves like a ideal low pass filter. The others are wrong.

Answer 221

1 Spectral leakage occurs when signals analyzed with a slightly truncated window don't align with the assumed periodicity. This mismatch causes discontinuities at window edges, spreading signal energy into neighboring frequency bins and distorting the spectrum. Spectral leakage isn't inherent but rather arises from specific conditions, not always occurring when using windows. Therefore, option 1 is incorrect as spectral leakage is contingent on the signal's alignment with the analysis window.

Answer 222

B A) Increasing the ripple in the passband of a Chebyshev filter decreases the roll-off rate, making it similar to a Butterworth filter: This statement is incorrect because increasing the ripple in a Chebyshev filter's passband actually increases the roll-off rate, not decreases it. B) A Chebyshev filter with 0% ripple in the passband becomes a Butterworth filter, which is known for its maximally flat frequency response: This is correct. Butterworth filters are characterized by a maximally flat response in the passband, meaning there is no ripple. When a Chebyshev filter is designed with 0% ripple, it essentially adopts the characteristics of a Butterworth filter, emphasizing a smooth passband without fluctuations. C) Butterworth filters achieve a faster roll-off than Chebyshev filters by introducing ripple in the passband: This statement is incorrect. D) Chebyshev filters can only achieve a faster roll-off than Butterworth filters when the passband ripple exceeds 1%: This statement is misleading.

Answer 223

C Autocorrelation quantifies the similarity between a signal and its own shifted version at various lags, shedding light on patterns inherent to the signal, such as periodicity. For a discrete signal x[n], the autocorrelation at a specific lag k, symbolized as rₓ[k], is calculated by summing the product of the signal values with their values at the k-lagged positions. For a lag of 1, this calculation is as follows: rₓ[1] = x[0]*x[1] + x[1]*x[2] Applying this to x[n] = {2,1,3}: rₓ[1] = (2 * 1) + (1 * 3) = 2 + 3 = 5

Answer 224

b a) Incorrect. The impulse response function describes the time-domain response of a filter to an impulse input, not the frequency response. b) Correct. c) Incorrect. The process of converting a time-domain signal into its frequency-domain representation is known as Fourier Transform, not the impulse response function. d) Incorrect. The cut-off frequency in low-pass filters is related to the filter's design and specifications, not directly to the impulse response function h(n)

Answer 225

a a) is correct because it accurately defines spectral leakage as the leakage of signal energy into adjacent frequency bins, leading to distortion in spectral analysis

Answer 226

B Ripple in filter design refers to the variations or fluctuations in amplitude observed within the passband or stopband of a filter. It is an important factor to consider in the filter design process, as excessive ripple may adversely affect the filter's performance in certain applications

Answer 227

A Windowing in signal processing multiplies a signal with a window function, which can cause abrupt transitions at the window's start and end. This discrepancy leads to spectral leakage, where energy spreads across different frequency components rather than remaining concentrated. This spreading effect is due to the abrupt changes, causing the energy to disperse into frequencies outside the main component, altering the signal's spectrum. Thus, "start" and "end" relate to window edges, and "spreading" describes the resulting dispersion of energy.

Answer 228

d Time domain filters are generally simpler to implement and require less computational resources compared to frequency domain filters. Frequency domain filters often introduce significant processing delays due to the need for Fourier transforms and inverse transforms. Time domain filters are effective in reducing noise and unwanted signal components in real-time data streams (e.g., time domain filters can adapt their parameters in real-time based on the characteristics of the input signal).

Answer 229

D Roll-off in digital filters refers to how quickly the filter attenuates frequencies beyond the passband, specifically, the steepness or slope of the filter's attenuation in the transition band between the passband and stopband. A steep roll-off means the filter more sharply reduces the amplitude of frequencies outside its passband, which is desirable in many applications for a clear distinction between filtered and attenuated frequencies. So the clear answer is D)

Answer 230

a A low-pass filter is designed to allow frequencies below a certain cutoff frequency to pass through unaffected, while attenuating frequencies above the cutoff. This means that signals with frequencies lower than the cutoff frequency will be transmitted with little or no attenuation, while signals with frequencies higher than the cutoff will be reduced in magnitude. Therefore, the correct choice is a) Low-pass filter.

Answer 231

B Spectral leakage, an artifact that occurs when a signal is not periodic within its sample window, can distort the signal's frequency spectrum. This distortion can be mitigated by applying window functions, such as the Hamming or Blackman window, to smooth the signal's abrupt ends before performing the DFT. This process reduces the leakage by minimizing the energy spread caused by the abrupt signal cutoff, enhancing the filter's effectiveness and the accuracy of the spectral analysis.

Answer 232

D Option A - The System's Response to a Loud Pop: A loud pop is a short, sharp sound that can somewhat mimic an impulse. Option B - The System's Response to a Clap: Like a loud pop, a clap is also a brief and high-energy sound. It approximates an impulse. Option C - The System's Response to a Song: A song is a complex signal with varying frequencies, rhythms, and intensities. It bears little resemblance to an impulse, which is a singular, instantaneous event. In the context of the convolution reverb experiment, we see that a loud pop and a clap were used to characterize the building(system).

Answer 233

3 A low-pass filter is designed to attenuate or block high-frequency components of a signal while allowing low-frequency components to pass through relatively unaffected. This means that high-frequency components, such as noise or interference, are suppressed by the filter, resulting in a smoother output signal with reduced high-frequency content. Therefore, answer is 3

Answer 234

C,D Abrupt changes(Discontinuous points) in boundary could cause Spectral leakage. Since hamming window reduces the side effects of truncation by smoothing, it can reduce spectral leakage. So (a) and (b) are correct. (c) If the window size is small, truncation can occur, leading to spectral leakage. (d) Spectral leakage is not directly caused by low sampling frequency. Even at high sampling frequencies, if the signal contains frequency components that do not align well with the window boundaries, spectral leakage can still occur. Therefore, (c) and (d) are incorrect

Answer 235

C Spectral leakage occurs when a signal is truncated, leading to the spreading of its energy across other frequencies. To mitigate this effect, the slide recommends smoothing the abrupt ends of the signal by applying windowing functions such as Hamming or Blackman. These windows reduce the leakage by tapering the signal at its ends, rather than abruptly cutting it off, which is characteristic of a rectangular window.

Answer 236

B Window functions like Hamming and Hanning are commonly used in signal processing to reduce spectral leakage by smoothing abrupt ends of the signal. In the frequency domain, the input signal and filter can create an output signal through multiplication, not convolution. If the windowing size is not sufficient to contain the original signal, spectral leakage can occur. Abrupt changes or ends in the signal can result in spectral leakage. Therefore option B is incorrect.

Answer 237

C The correct answer is C. This choice is optimal because it covers the frequency range that corresponds to the typical human heart rate, which, when converted to Hertz, is 1 Hz to 3 Hz. By selecting a Band Pass Filter with a range from 0.5Hz to 4Hz, we ensure that the heart rate signals are effectively isolated from the noises happening at 10 Hz, 40 Hz, and 0.2 Hz. This range not only captures the relevant heart rate frequencies but also provides a margin to account for potential outliers.

Answer 238

B The Butterworth filter is known for its maximally flat frequency response in the passband, meaning it has no ripples. This design choice results in a gradual transition or slope as it approaches the cutoff frequency, moving towards the stopband. The slope, or roll-off, is a byproduct of attempting to maintain this flatness as closely as possible up to the cutoff frequency, beyond which attenuation begins. This gradual transition helps to reduce distortions within the passband while effectively attenuating unwanted frequencies in the stopband.

Answer 239

C To reduce passband ripple in an FIR filter, the most effective strategy is utilizing a windowing technique (C). Windowing smooths out the filter's impulse response, reducing side lobes and effectively diminishing ripple in both the passband and stopband. On the other hand, zero-padding (A) mainly enhances frequency resolution in spectral analysis without directly affecting the ripple. Decreasing the filter order (B) might simplify the filter but can increase ripple due to less precise control over the filter's frequency response. Narrowing the frequency band (D) does not directly influence passband ripple.

Answer 240

B FIR (Finite Impulse Response) filters can maintain a linear phase response, which is crucial for applications where preserving the wave shape of the original signal is important, as it ensures all frequency components of the signal are delayed by the same amount when passing through the filter. This property makes FIR filters particularly suitable for phase-sensitive applications. On the other hand, IIR (Infinite Impulse Response) filters, due to their recursive nature, can introduce phase distortion, as they may not delay all frequency components uniformly. This difference highlights the importance of choosing the right filter type based on the specific requirements of a signal processing task.

Answer 241

B Spectral leakage is a phenomenon that occurs when a signal is windowed, leading to the spreading of energy from the (mainlobe to the sidelobes. This effect reduces the spectral resolution because energy that should be concentrated in the main frequency band leaks into other frequencies. It is prompting the use of various windowing techniques to minimize the leakage and preserve as much spectral accuracy as possible.

Answer 242

B Option C) is the correct answer as it accurately interprets the frequency spectrum of a real signal, indicating the presence of fundamental and harmonic frequencies. These peaks are characteristic of periodic mechanical sources, such as the blades of a submarine's propeller, which generate distinct peaks at these frequencies due to their rotation.

Answer 243

C Increasing the order of a Butterworth filter makes its transition from passband to stopband steeper, allowing for a clearer distinction between allowed and blocked frequencies. This improves the filter's precision in separating desired signals from noise.

Answer 244

B A, C, D are all correct explanations about spectral leakage. The main reason for spectral leakage is the inappropriate window size during sampling which in turn leads to energy leaking out from the mainlobe to the sidelobes after DFT. One way to reduce this is to smooth out the abrupt ends of the windowed signal. B says that high autocorrelation leads to spectral leakage which is not the case.

Answer 245

B B is the correct answer according to the slides. For large N, the function behaves like a ideal low pass filter. The others are wrong.

Answer 246

1 Spectral leakage occurs when signals analyzed with a slightly truncated window don't align with the assumed periodicity. This mismatch causes discontinuities at window edges, spreading signal energy into neighboring frequency bins and distorting the spectrum. Spectral leakage isn't inherent but rather arises from specific conditions, not always occurring when using windows. Therefore, option 1 is incorrect as spectral leakage is contingent on the signal's alignment with the analysis window.

Answer 247

B A) Increasing the ripple in the passband of a Chebyshev filter decreases the roll-off rate, making it similar to a Butterworth filter: This statement is incorrect because increasing the ripple in a Chebyshev filter's passband actually increases the roll-off rate, not decreases it. B) A Chebyshev filter with 0% ripple in the passband becomes a Butterworth filter, which is known for its maximally flat frequency response: This is correct. Butterworth filters are characterized by a maximally flat response in the passband, meaning there is no ripple. When a Chebyshev filter is designed with 0% ripple, it essentially adopts the characteristics of a Butterworth filter, emphasizing a smooth passband without fluctuations. C) Butterworth filters achieve a faster roll-off than Chebyshev filters by introducing ripple in the passband: This statement is incorrect. D) Chebyshev filters can only achieve a faster roll-off than Butterworth filters when the passband ripple exceeds 1%: This statement is misleading.

Answer 248

C Autocorrelation quantifies the similarity between a signal and its own shifted version at various lags, shedding light on patterns inherent to the signal, such as periodicity. For a discrete signal x[n], the autocorrelation at a specific lag k, symbolized as rₓ[k], is calculated by summing the product of the signal values with their values at the k-lagged positions. For a lag of 1, this calculation is as follows: rₓ[1] = x[0]*x[1] + x[1]*x[2] Applying this to x[n] = {2,1,3}: rₓ[1] = (2 * 1) + (1 * 3) = 2 + 3 = 5

Answer 249

b a) Incorrect. The impulse response function describes the time-domain response of a filter to an impulse input, not the frequency response. b) Correct. c) Incorrect. The process of converting a time-domain signal into its frequency-domain representation is known as Fourier Transform, not the impulse response function. d) Incorrect. The cut-off frequency in low-pass filters is related to the filter's design and specifications, not directly to the impulse response function h(n)

Answer 250

a a) is correct because it accurately defines spectral leakage as the leakage of signal energy into adjacent frequency bins, leading to distortion in spectral analysis

Answer 251

B Ripple in filter design refers to the variations or fluctuations in amplitude observed within the passband or stopband of a filter. It is an important factor to consider in the filter design process, as excessive ripple may adversely affect the filter's performance in certain applications