1. Introduction to Artificial Intelligence Flashcards
Course 1 / 16 (291 cards)
1
Q
What is AI?
A
- Anything that simulates human intelligence through computer systems,
- Utilises algorithms and data to function,
- Enables machines to perform tasks requiring human intelligence
2
Q
How should you think of AI?
A
AI should be thought of as augmented intelligence. AI should not replace human experts but merely extend their capabilities by doing things that humans or machines could not do on their own.
3
Q
What is innate intelligence?
A
Innate intelligence is the intelligence that governs every activity in our bodies. For example, this intelligence is what causes an oak tree to grow out of a little seed.
4
Q
How does AI learn?
A
Machines are provided with the ability to examine and create machine learning models.
This is done in various ways such as supervised learning, unsupervised learning and reinforcement learning.
5
Q
What can AI be described by?
A
Strength, breadth, and application.
6
Q
What is weak or narrow AI?
A
This is AI that is applied to a specific domain. For example, language translators, self-driving cars, and recommendation engines.
Applied AI can perform specific tasks, but not learn new ones, making decisions based on programmed algorithms and training data.
7
Q
What is strong or generalised AI?
A
This is AI that can interact and operate a wide variety of independent and unrelated tasks.
It can learn new tasks to solve new problems, and it does this by teaching itself new strategies.
Generalised intelligence is the combination of many AI strategies that learn from experience and can perform at a human level of intelligence.
8
Q
What is Super AI or Conscious AI?
A
This is AI with human-level consciousness, which would require it to be self-aware.
Because we are not yet able to adequately define what consciousness is, it is unlikely we will be able to create Super AI in the near future.
9
Q
How is AI multi-disciplinary?
A
- Computer Science and Electrical Engineering determine how AI is implemented in software and hardware.
- Mathematics and Statistics determine viable models and measure performance.
- Because AI is modelled on how the human brain works, psychology and linguistics play an essential role in understanding how AI might work.
- Philosophy provides guidance on ethical considerations and intelligence.
10
Q
Artificial Intelligence vs. Generative AI (GenAI)
A
Artificial Intelligence is an augmented intelligence that helps experts scale their capabilities while machines handle time-consuming tasks such as recognising speech, playing games, and making decisions.
Generative AI, on the other hand, is an AI technique that can generate new and unique data ranging from images and music to texts and virtual worlds.
11
Q
How does GenAI differ from conventional AI?
A
Conventional AI relies on predefined rules and patterns while GenAI uses deep learning techniques and relies on large datasets to generate new data.
12
Q
Generative AI and LLM
A
Large Language Model (LLM) is a type of AI that uses deep learning techniques to process and generate natural language.
Generative AI can develop new and powerful LLM algorithms or architectures.
Generative AI can also incorporate LLM into a larger, more advanced AI system.
13
Q
What are the benefits of Generative AI?
A
- Creativity and Innovation
- Cost and time savings
- Personalisation
- Scalability
- Robustness
- Exploration of new possibilities
14
Q
What are some use cases for GenAI?
A
- Healthcare and precision medicine
- GenAI can identify genetic mutations
- Provide personalised treatment options
- Help doctors practice procedures and develop treatments
- Agriculture
- GenAI can optimise crop yields
- Develop new, more resistant crop varieties
- Biotechnology
- Generative AI can aid in the development of new drugs and therapies by:
- identifying potential drug targets
- simulating drug interactions
- forecasting drug efficacy
- Generative AI can aid in the development of new drugs and therapies by:
- Forensics
- Generative AI can help solve crimes by:
- analysing DNA evidence
- identifying suspects
- Generative AI can help solve crimes by:
- Environmental conservation
- Generative AI can support the protection of endangered species by:
- analysing genetic data and suggesting breeding and conservation strategies
- Generative AI can support the protection of endangered species by:
- Creative
- Generative AI can produce unique digital art, music, and video content for:
- advertising and marketing campaigns
- and generate soundtracks for films and video games
- Generative AI can produce unique digital art, music, and video content for:
- Gaming
- Generative AI can create interactive game worlds and
- generate new characters, levels, and objects in real-time
15
Q
How does AI mean different things to different people?
A
For a video game designer, AI means affecting the way bots play or how the environment adapts to the player; whereas for a data scientist, AI is a way of exploring and classifying data to meet specific goals.
16
Q
What is the reason we are able to talk to virtual assistants such as Alexa and Siri?
A
It is because of AI algorithms that learn by example. The natural language processing and natural language generation capabilities that come with AI aid our interactions with these virtual assistants so they can talk back to us.
17
Q
How does Generative AI work?
A
It works by leveraging machine learning and deep learning techniques to learn patterns and generate original content
18
Q
What does Generative AI do?
A
It enables applications to create, generate, and simulate new content
19
Q
Name and describe different types of AI
A
The different types of AI include:
Diagnostic/descriptive AI: Focuses on assessing the correctness of behavior by analyzing historical data to understand what happened and why.
Predictive AI: Concerned with forecasting future outcomes based on historical and current data.
Prescriptive AI: Focuses on determining the optimal course of action by providing recommendations based on data analysis.
Generative/cognitive AI: Involved in producing various types of content, such as code, articles, images, and more.
Reactive AI: Designed to respond to specific inputs with predetermined responses.
Limited memory AI: Have the ability to use past experiences to inform current decisions.
Theory of Mind AI: Advanced type of AI that aims to understand human emotions, beliefs, and intentions.
Self-aware AI: Represents the most advanced form of AI, which has its own consciousness and self-awareness.
Narrow AI (Weak AI): Designed to perform a specific task or a limited range of tasks.
General AI (Strong AI): Can understand, learn, and apply knowledge across a wide range of tasks like human intelligence.
20
Q
Diagnostic/descriptive AI
A
Diagnostic/descriptive AI
Diagnostic or descriptive AI focuses on assessing the correctness of behavior by analyzing historical data to understand what happened and why. This type of AI is instrumental in identifying patterns and trends, performing comparative analyses, and conducting root cause analyses.
Capabilities:
Scenario planning: Helps in creating different future scenarios based on historical data.
Pattern/trends recognition: Identifies recurring patterns and trends within data sets.
Comparative analysis: Compares various data points to find correlations and insights.
Root cause analysis: Determines the underlying reasons behind specific outcomes.
21
Q
Predictive AI
A
Predictive AI
Predictive AI is concerned with forecasting future outcomes based on historical and current data. This type of AI is used extensively in predicting customer behavior, market trends, and other forward-looking insights.
Capabilities:
Forecasting: Predicts future trends and events.
Clustering and classification: Groups similar data points and classifies them into predefined categories.
Propensity model: Assesses the likelihood of specific outcomes based on current data.
Decision trees: Utilize a tree-like model of decisions to predict outcomes.
22
Q
Prescriptive AI
A
Prescriptive AI
Prescriptive AI focuses on determining the optimal course of action by providing recommendations based on data analysis. It goes beyond prediction by suggesting actions that can help achieve desired outcomes.
Capabilities:
Personalization: Tailors recommendations and experiences to individual needs.
Optimization: Identifies the most efficient ways to achieve goals.
Fraud prevention: Detects and prevents fraudulent activities through analysis.
Next best action recommendation: Provides actionable insights on the next steps to take.
23
Q
Generative/cognitive AI
A
Generative/cognitive AI
Generative or cognitive AI is involved in producing various types of content, such as code, articles, images, and more. This type of AI mimics human creativity and cognitive processes to automate and assist in content creation.
Capabilities:
Advises: Offers expert advice and recommendations.
Creates: Produces new content, such as text, images, and code.
Protects: Enhances security measures through intelligent analysis.
Assists: Provides assistance in various tasks, improving efficiency.
Automates: Automates repetitive tasks to save time and resources.
24
Q
Reactive AI
A
Reactive AI
Reactive AI systems are designed to respond to specific inputs with predetermined responses. They do not have memory or the ability to learn from past experiences, making them suitable for tasks that require immediate reactions.
Capabilities:
Rule-based actions: Executes specific actions based on predefined rules.
Instant responses: Provides immediate responses to inputs.
Static data analysis: Analyzes current data without considering past interactions.
25
Limited memory AI
Limited memory AI:
Limited memory AI systems have the ability to use past experiences to inform current decisions. They can learn from historical data to improve their performance over time. This type of AI is commonly used in autonomous vehicles and recommendation systems.
Capabilities:
Learning from data: Uses historical data to make informed decisions.
Pattern recognition: Identifies patterns over time to improve accuracy.
Adaptive responses: Adapts responses based on previous interactions.
26
Theory of Mind AI
Theory of Mind AI:
Theory of Mind AI is an advanced type of AI that aims to understand human emotions, beliefs, and intentions. It is still in the research stage and seeks to interact more naturally with humans by comprehending their mental states.
Capabilities:
Emotion recognition: Identifies and responds to human emotions.
Social interaction: Engages in more natural and human-like interactions.
Intent prediction: Predicts human intentions based on context and behavior.
27
Self-aware AI
Self-aware AI:
Self-aware AI represents the most advanced form of AI, which has its own consciousness and self-awareness. This type of AI can understand and react to its own emotions and states. It remains a theoretical concept and has not yet been realized.
Capabilities:
Self-diagnosis: Evaluates its own performance and health.
Autonomous learning: Learns independently without human intervention.
Adaptive behavior: Adjusts behavior based on self-awareness.
28
Narrow AI (Weak AI)
Narrow AI (Weak AI):
Narrow AI is designed to perform a specific task or a limited range of tasks. It excels in a single area but lacks generalization capabilities. Most current AI applications fall under this category.
Capabilities:
Task specialization: Excels in performing specific tasks.
High accuracy: Achieves high performance in its designated area.
Efficiency: Operates efficiently within its scope of specialization.
29
General AI (Strong AI)
General AI (Strong AI):
General AI, like human intelligence, can understand, learn, and apply knowledge across a wide range of tasks. It can also transfer knowledge from one domain to another and adapt to new situations autonomously.
Capabilities:
Cross-domain learning: Applies knowledge across various domains.
Autonomous decision making: Makes decisions independently in diverse scenarios.
Human-like understanding: Understands and processes information similar to humans.
30
What are some known applications of GenAI?
- GPT-4
- ChatGPT
- Bard
- GitHub CoPilot
31
How do cognitive systems interpret the information they read?
Cognitive systems use processes similar to the decision-making process of humans to interpret and generate hypotheses about the information they read.
32
What do cognitive systems rely on to understand the intent and context of a user's language?
Cognitive systems rely on natural language governed by rules of grammar, context, and culture.
33
How do cognitive computing systems differ from conventional computing systems?
They differ in that they can:
- Read and interpret unstructured data, understanding not just the meaning of words but also the intent and context in which they are used.
- Reason about problems in a way that humans reason and make decisions.
- Learn over time from their interactions with humans and keep getting smarter.
34
What are the differences between artificial intelligence, machine learning, deep learning and neural networks?
- artificial intelligence is a branch of computer science dealing with the simulation of intelligent behaviour
- machine learning is a subset of AI that uses computer algorithms to analyse data and make intelligent decisions based on what it has learned, without being explicitly programmed
- deep learning is a subset of machine learning that uses layered neural networks to simulate human decision-making
- neural networks in AI are a small collection of computing units (neurons) that take incoming data and learn to make decisions over time
35
What behaviours do AI systems seek to demonstrate?
| People Love Reading Poems, Kids Pref Magic Marvels, Cats Sing Incredibly
They seek to demonstrate behaviours associated with human intelligence such as:
- planning
- learning
- reasoning
- problem-solving
- knowledge
- perception
- motion
- manipulation
- creativity
- social intelligence
36
What are some characteristics of machine learning algorithms
- machine learning algorithms are trained with large sets of data
- they learn from examples
- they do not follow rules-based algorithms
37
What enables machines to solve problems on their own?
Machine learning
38
What can deep learning algorithms do?
They can label and categorise information and identify patterns.
39
How does machine learning differ from traditional programming?
- In traditional programming, you have data and create algorithms in order to find answers.
- this algorithm will not change
- In machine learning, you already have the data and the answers and are using that information to create an algorithm.
- What you get at the end is a set of rules that determine what the machine learning model will be.
40
What does machine learning rely on?
Defining behavioural rules by examining and comparing large data sets to find common patterns.
41
What is supervised learning?
A type of machine learning where an algorithm is trained on human-labeled data. The more data you provide a supervised learning algorithm, the more precise it becomes in classifying new data.
42
What is unsupervised learning?
A type of machine learning that relies on giving the algorithm unlabelled data and letting it find patterns by itself. You provide the input, but not labels, and let the machine infer qualities.
43
What is reinforcement learning?
A type of machine learning that relies on providing a machine learning algorithm with a set of rules and constraints and letting it learn how to achieve its goals. You define the state, the desired goal, allowed actions, and constraints.
44
What is a machine learning model?
The algorithm used to find patterns in the data without the programmer having to explicitly program these patterns.
45
What are the three categories of supervised learning?
- Regression
- Classification
- Neural networks
46
How are regression models built?
By looking at the features - X and the result - Y, where Y is a continuous variable.
Essentially, regression estimates continuous values.
47
What do neural networks refer to?
Structures that imitate the structure of the human brain.
48
What is classification?
The process of predicting the class of given data points.
49
What does classification focus on?
It focuses on discrete values it identifies.
We can define discrete class labels - Y based on many input features - X.
50
What are some forms of classification?
- Decision trees
- Support vector machines
- Logistics regression
- Random forests
51
What are features?
Distinctive properties of input patterns that help in determining the output categories or classes of output.
52
What is the meaning of training, in an ML context?
It refers to using a learning algorithm to determine and develop the parameters of your model.
53
What do you typically do with data sets in machine learning?
You split them into 3 sets:
1. Training set
2. Validation set
3. Test set
54
What is the training set for?
It is for training the algorithm
55
What is the validation set for?
It is used to validate results and fine-tune the algorithm's parameters
56
What is the test set for?
It is used to evaluate how good the model works on unseen data.
57
What does deep learning do?
It layers algorithms to create a neural network, an artificial representation of the structure and functionality of the brain
58
What does deep learning enable AI systems to do?
Continuously learn on the job and improve the quality and accuracy of results.
59
How are deep learning algorithms developed?
- Developers and engineers configure the number of layers and the type of functions that connect the outputs of each layer to the inputs of the next
60
Which older generation issue is fixed by deep learning, and how?
The efficiency and performance of older-generation machine learning algorithms plateau as the datasets grow. Deep learning algorithms plateau continue to improve as they are fed more data.
61
Which tasks has deep learning proved to be very efficient at?
- Image captioning
- Voice recognition
- Facial recognition
- Medical Imaging
62
What is the name of the process that neural networks use to learn?
Backpropagation.
63
How does backpropagation work?
- backpropagation uses a set of training data that match known inputs to desired outputs
- first, the inputs are plugged into the network, and outputs are determined
- then, an error function determines how far the given output is from the desired output
- finally, adjustments are made in order to reduce errors
64
What is a collection of neurons called?
A layer
65
What do layers do?
They take in an input and provide an output.
66
What do neural networks always have?
One input layer and one output layer. They will also have more than one hidden layer which simulates the type of activity that goes on in the human brain.
67
What do hidden layers do?
They take in a set of weighted inputs and produce an output through an activation function.
68
What do you call a neural network having more than one hidden layer?
A deep neural network.
69
What are the simplest and oldest types of neural networks?
Perceptrons.
70
What are perceptrons?
They are single-layered neural networks consisting of input nodes connected directly to an output node.
Input layers forward the input values to the next layer, by means of multiplying by a weight and summing the result.
71
What property to input and output nodes have?
Bias
72
What is bias?
A special type of weight that applies to a node after the other inputs are considered.
73
What determines how a node responds to its inputs?
An activation function.
74
How does an activation function work?
The function is run against the sum of the inputs and bias, and then the result is forwarded as an input.
75
What is a critical component to the success of a neural network?
The type of activation function chosen.
76
What are Convolutional Neural Networks (CNNs)?
They are multi-layered neural networks that take inspiration from the animal visual cortex.
77
Where are CNNs useful?
They are useful in applications such as image processing, video recognition, and natural language processing.
78
What is a convolution?
It is a mathematical operation where a function is applied to another function and the result is a mixture of the two functions.
79
What are convolutions good at?
They are good at detecting simple structures in an image and putting those simple features together to construct more complex features.
80
What are Recurrent Neural Networks?
Recurrent Neural Networks or RNNs are multi-layered neural networks that perform the same task for every element of a sequence, with prior outputs feeding subsequent stage inputs.
81
How are Recurrent Neural Networks recurrent?
They are recurrent because they perform the same task for every element of a sequence, with prior outputs feeding subsequent stage inputs.
82
What is the unique characteristic of Recurrent Neural Networks (RNNs) in handling sequence data?
RNNs make use of information in long sequences, each layer of the network representing the observation at a certain time.
83
How is input processed in a general neural network?
- an input is processed through a number of layers
- and an output is produced with an assumption that the two successive inputs are independent of each other, but that may not hold true in certain scenarios.
84
Give a scenario where a recurrent neural network would work better than a general neural network
- In a general neural network, output is produced with an assumption that the two successive inputs are independent of each other, but that may not hold true in certain scenarios
- For example, when we need to consider the context in which a word has been spoken, in such scenarios, dependence on previous observations has to be considered to produce the output.
- RNNs can make use of information in long sequences, each layer of the network representing the observation at a certain time.
85
What are the most common areas of AI?
- Natural Language Processing
- Speech
- Computer Vision
86
What is Natural Language Processing?
Natural Language Processing (NLP) is a subset of artificial intelligence that enables computers to understand the meaning of human language, including the intent and context of use.
87
What does speech-to-text enable machines to do?
It enables them to convert speech to text by identifying common patterns in the different pronunciations of a word, mapping new voice samples to corresponding words.
88
What does speech synthesis enable machines to do?
It enables machines to create natural-sounding voice models, including the voices of particular individuals.
89
What does Computer Vision enable machines to do?
It enables machines to identify and differentiate objects in images the same way humans do.
90
What AI application areas do self-driving cars use?
Self-driving cars utilise NLP, speech, and most importantly, computer vision.
91
What is the hot topic in AI?
How to use it responsibly
92
What are the five pillars of AI ethics?
- Explainability
- Transparency
- Robustness
- Privacy
- Fairness
93
What is important to remember about AI ethics?
That they are not a "one-and-done" activity. These ethics need to be considered throughout the entire AI lifecycle.
94
What does facial recognition do?
It simply detects a human face in an image.
95
What does facial authentication do?
It provides a one-to-one authentication matching a single face image to another single face image.
96
What does facial matching do?
It provides a one-to-many matching to uniquely identify an individual from a database of images.
97
What is an example of an AI use case that raises ethical concern?
Using facial recognition technology to scan individuals in crowds without consent
98
What raises AI ethical concerns in the workforce?
Bias could permeate AI in decision-making. It could make discriminatory decisions based on race, age, and gender, for example if care isn't taken.
99
What is an ethical concern about AI on social media platforms?
Marketing on social media is dependent on data collected from users on those platforms - but its not always explicit what data has been collected or if users have even provided consent to have their data collected.
100
What makes AI ethics a socio-technical challenge?
AI is pervasive in all aspects of our daily lives, whether we realise it or not.
For example, we encounter AI every time we buy something using a credit card online, when we search something on the web, or when we post, like, or follow somebody on a social platform.
101
Besides technical tools, what else do you need to build trustworthy AI?
You need to utilise principles, guardrails, well-defined processes, and governance.
102
What does it mean to build an use AI ethically?
It means embedding AI ethics into the design and development of the AI lifecycle.
103
What are IBM's three guiding principles for AI?
1. The purpose of AI is to augment - not replace - human intelligence
2. Data and insights belong to their creator
3. New technology, including AI, must be transparent and explainable
104
What are the five pillars supporting IBM's guiding principles?
1. Transparency
2. Explainability
3. Fairness
4. Robustness
5. Privacy
105
How can organisations put AI ethics into action?
- The first step to building AI ethics is developing awareness and understanding
- The second step is to put into place a governance structure that works at scale
- The third step is operationalisation
- It is critical that every contributor in the organisation knows what they need to do to support AI ethics
106
What does the operationalisation of AI ethics mean?
It means having clarity around the principles and pillars of AI ethics.
107
What is AI ethics?
It is a multidisciplinary field that investigates how to maximise AI's beneficial impacts while reducing risks and adverse impacts.
108
When is AI explainable?
When it can show how and why it arrived at a particular outcome or recommendation.
109
When is AI fair?
When it treats individuals or groups equitably.
110
When is AI robust?
When it can effectively handle exceptional conditions, like abnormal input or adversarial attacks.
111
When is AI transparent?
When appropriate information is shared with humans about how the AI system was designed and developed.
112
When is AI private?
When it is designed to prioritise and safeguard humans' privacy and data rights.
113
What does bias in AI do?
It gives systematic advantages to certain groups or individuals
114
How can bias emerge in AI?
- AI is trained using decisions that human decision-makers have mad in the past
- These human decision-makers may have been implicitly or explicitly biased, and those biases may then may reflected in the training data
- The sampling of the data could also introduce bias (for example, through overrepresentation or underrepresentation)
- Bias can be introduced through the data processing stage
- The way that a problem is conceptualised can also introduce bias
115
What are some high-level actions that can help mitigate bias in AI?
- The first step is assembling diverse teams
- Search for high-quality data sets
- There are many possible technical approaches to mitigating bias
116
What is a regulation?
A government rule enforceable by law
117
What is precision regulation?
Regulation that aims to be risk-based, context-specific, and allocates responsibility to the part closest to the risk, which might shift throughout the AI lifecycle
118
What are IBM's five policy imperatives for organisations that provide or own AI systems?
1. Designate a lead AI ethics official
2. Develop different rules for different risks
3. Don't hide your AI - make it transparent
4. Explain your AI
5. Test your AI for bias
119
What is AI governance?
This is an organisation's act of governing AI through its corporate instructions, staff, processes, and systems
120
What is the objective of governance?
To deliver trustworthy AI by establishing requirements for accountability, responsibility, and oversight.
121
What are the benefits of governance?
- Trust
- Efficiency
- Compliance
122
What does a successful governance program account for?
People, processes, and tools
123
What does ESG stand for?
Environmental, social, and governance
124
What are the 3 pillars of the impact of ESG?
1. Environmental impact
2. Equitable impact
3. Ethical impact
125
What does the governance aspect of ESG pertain to?
It pertains to creating innovations, policies, and practices that prioritise ethics, trust, transparency, and above all, accountability.
126
What is driving the industry focus on trustworthy AI?
Corporate social responsibility posture, concerns around reputational risk and a growing set of regulations.
127
What are the stages in the AI lifecycle?
- Scope and plan
- Collect and organize
- Build and test
- Validate and deploy
- Monitor and manage
128
Why should you apply Enterprise Design Thinking principles during the scoping and planning stage of the AI lifecycle?
Design Thinking helps identify and define various business and technical aspects of bias/fairness, robustness, explainability etc. in the context of the use case.
129
What is the 'Scope and Plan' stage of the AI lifecycle?
This stage guides the prioritisation of use cases and the development of an AI action plan.
130
What is the 'Collect and Organise' stage of the AI lifecycle for?
This stage allows the data consumer to find and access relevant datasets.
131
What happens during the 'Build and Test' stage of the AI lifecycle?
Data science teams explore and prepare the data, and build, train and test their AI/ML models during this stage. The activities during this stage are best undertaken as a set of agile sprints.
132
What is an important activity that needs to be done during the 'Build and Test' stage of the AI lifecycle?
- It is important that bias in data be checked at this stage even before any model building work starts.
- This is an inner guardrail for fairness, and such guardrails can be put in place during and after the model building steps.
- Model robustness, explainability and other aspects can be similarly accounted for and tested during this stage.
133
What happens during the 'Validate and Deploy' stage of the AI lifecycle?
- This stage involves validation of the model and deployment into production.
- The team validates quality, robustness, fairness, etc. and generates validation reports. It is important to capture the validation results for reference and comparison.
134
What happens if a model passes validation?
The Ops team can promote it to the production environment using an MLOps pipeline. The model is deployed there, either for online or batch invocation.
135
What happens during the 'Monitor and Manage' stage of the AI lifecycle?
- In this stage, the Ops team sets up ongoing monitoring and management of the AI/ML model in production.
- The team configures monitors for periodic scheduled collection of metrics.
- Quality, robustness, and fairness are monitored at frequencies dictated by business needs.
136
What are the 3 pillars of the AI governance policy framework?
- Accountability
- proportionate to the risk profile of the application and the role of the entity providing, developing, or operating an AI system to control and mitigate unintended or harmful outcomes for consumers.
- Transparency
- in where the technology is deployed, how it is used, and why it provides certain determinations.
- Fairness and security
- validated by testing for bias before AI is deployed and re-tested as appropriate throughout its use, especially in automated determinations and high-risk applications.
137
What do the OECD AI Principles suggest would help create a solid accountability bedrock for the AI Governance Framework?
Promoting a policy environment that supports an agile transition from the research and development stage to the deployment and operation stage for trustworthy AI systems.
138
How can differentiating accountability mitigate the harm caused by AI?
It can do this by directing resources and oversight to specific applications of AI based on the severity and likelihood of potential harms arising from the end-use and user of such systems.
139
What is the ideal scenario for AI in the modern world?
It's not to try and develop a system that completely autonomously handles every aspect of a problem but to have a collaboration between machines doing what they do best and humans doing what they do best.
140
What is likely to be the next big growth area for AI?
Understanding and generating natural language is likely to be the next big growth area for AI, along with vision systems for the blind.
141
Why are a large number of businesses unable to leverage the capabilities of AI?
A large proportion of their data is locked in silos and not business-ready. There is no AI without IA.
142
What does IA mean?
Information Architecture
143
What do businesses use the AI Ladder for?
It is used for the successful implementation of AI
144
What are the four steps that make up the AI Ladder?
1. Collect: Make data simple and accessible.
1. Collect data of every type regardless of where it resides, enabling flexibility in the face of ever-changing data sources.
2. Organise: Create a business-ready analytics foundation.
1. Organise all data into a trusted, business-ready foundation with built-in governance, protection, and compliance.
3. Analyse: Build and scale AI with trust and transparency.
1. Analyse data in smarter ways and benefit from AI models that empower organisations to gain new insights and make better, smarter decisions.
4. Infuse: Operationalise AI throughout the business.
1. Apply AI across the enterprise in multiple departments and within various processes—drawing on predictions, automation, and optimisation.
145
What is an important step before the AI Ladder?
Modernising your data and making it ready for an AI and hybrid cloud world.
146
What is some helpful advice for AI?
- Because the technology is moving so fast, it is best not to get overly attached to a particular technology, technique, or implementation
- Apply what you learn
- Spend time learning maths and science
- Know how to use software APIs and understand the question being asked
147
Generative AI model architectures
Generative AI model architectures include VAEs, GANs, autoregressive models, and Transformers.
148
Variational autoencoders (VAEs)
Variational autoencoders (VAEs):
Encoder: Input data into a latent space representation
Latent space captures essential data characteristics
Decoder: Generates outputs based on this representation
149
Generative adversarial networks (GANs)
Generative adversarial networks (GANs):
Generator: New data samples
Discriminator: Verify the generated data
150
Autoregressive models
Autoregressive models:
Create data sequentially
Consider the context
151
Transformers
Transformers:
Generate text sequences
Perform cross-language translations effectively
152
What is machine learning?
Machine learning is a subset of AI that uses algorithms to analyze data, make decisions without explicit programming, and enable autonomous problem-solving.
153
What are the three main types of machine learning?
Supervised learning
Unsupervised learning
Reinforcement learning
154
What is supervised learning?
A type of machine learning where the algorithm is trained on labeled data to classify new data. It becomes more precise with more data.
155
Name two categories of supervised learning and what they do.
Regression: Estimates continuous values.
Classification: Focuses on discrete values.
156
What is unsupervised learning?
A type of machine learning that finds patterns in unlabeled data. It is useful for clustering similar data points and detecting anomalies.
157
What is reinforcement learning?
A type of machine learning where an algorithm achieves goals within a set of rules and constraints by maximizing rewards. It is useful for tasks like playing chess or navigating.
158
What are the three sets used to train a machine learning model?
Training set: Trains the algorithm
Validation set: Fine-tunes and validates the model
Test set: Evaluates the model's performance
159
What is deep learning?
Deep learning uses neural networks with multiple layers to analyze complex data. It allows continuous improvement and learning and enhances AI's natural language understanding.
160
Give six examples of tasks that deep learning excels at.
Image captioning
Voice recognition
Facial recognition
Medical imaging
Language translation
Driverless cars
161
Describe the structure of a neural network.
A neural network is a computational model consisting of interconnected nodes with three layers: an input layer, one or more hidden layers, and an output layer.
162
Name six types of neural networks.
Perceptron
Feed-forward
Deep feed-forward
Modular
Convolutional neural network
Recurrent neural networks
163
What are four generative AI model architectures, and what do they do?
VAEs (Variational autoencoders): Encode input data into a latent space representation, capture essential data characteristics, and generate outputs based on this representation.
GANs (Generative adversarial networks): Use a generator to create new data samples and a discriminator to verify the generated data.
Autoregressive models: Create data sequentially, considering the context.
Transformers: Generate text sequences and perform cross-language translations effectively.
164
What is the difference between unimodal and multimodal models?
Unimodal models process inputs and generate outputs within the same modality (like text-to-text). Multimodal models handle inputs from one modality and produce outputs in a different modality (like image-to-text).
165
What is cognitive computing?
Cognitive computing technology mimics human cognitive processes like thinking, reasoning, and problem-solving.
166
What is natural language processing (NLP)?
NLP aids computers in interpreting and producing human language. It uses machine learning and deep learning algorithms to understand a word's semantic meaning.
167
What is the difference between STT and TTS technology?
STT (Speech-to-text): Transforms spoken words into written text.
TTS (Text-to-speech): Converts written text into spoken words.
168
What is computer vision?
Computer vision enables machines to understand visual data by analyzing images or videos, drawing meaningful insights, and making informed decisions.
169
What are IoT devices?
IoT devices are a network of physical devices connected to the internet that collect and share data for processing and analysis.
170
What is cloud computing?
Cloud computing allows you to store and use data and services over the internet.
171
What is edge computing?
Edge computing refers to the practice of processing data closer to the source of generation, rather than relying on a centralized data center.
172
Give four examples of a real-world application that combines AI, IoT, cloud computing, and edge computing.
AI-powered traffic lights
Smart public transportation
Smart agriculture
Smart buildings
173
Retrieval-augmented generation (RAG)
RAG is a hybrid approach that combines generative models with retrieval-based methods. RAG works through a three-step process. First, given a query, the model retrieves relevant documents or pieces of information from a predefined corpus or database. Next, these retrieved documents are used to augment the input to the generative model, providing it with additional context. Finally, the generative model uses both the original query and the retrieved information to generate a response, ensuring that the output is both contextually rich and factually grounded. This process enhances the accuracy and relevance of the generated content, addressing many limitations of standalone generative models.
174
RAG: Key components
Retrieval component:
Function: The retrieval component of RAG is responsible for searching and extracting relevant information from a large corpus of documents or a knowledge base. This step ensures that the model has access to factual and contextually appropriate information.
Mechanism: Typically, this involves using a retriever model like BM25 or dense retrievers based on neural networks to find the most relevant passages or documents that match a given query.
Generation component:
Function: The generation component takes the retrieved information and uses it to generate coherent and contextually appropriate responses or text. This is achieved using a generative model like GPT-3 or BERT.
Mechanism: The generative model leverages the context provided by the retrieved documents to produce more accurate and relevant outputs, blending retrieval results with its generative capabilities.
175
Benefits of RAG
RAG improves the accuracy of responses by grounding generated text in actual data, enhancing factual correctness. It also boosts contextual relevance by incorporating real-time information retrieval, ensuring that responses are pertinent to the query. RAG's flexibility allows it to be adapted for various NLP tasks, making it a versatile tool in the AI toolkit. Additionally, its ability to retrieve the latest information ensures that responses are dynamically updated and current.
176
Applications of RAG
Question-answering systems: RAG can retrieve relevant documents to answer complex questions accurately.
Content creation: RAG assists in generating detailed and informative content for articles, reports, or creative writing.
Customer support: RAG can provide accurate and contextually relevant answers by retrieving the latest information from knowledge bases.
Search engines: RAG improves search results by providing detailed and accurate answers based on retrieved documents.
177
How do Copy.ai, Jasper, and Synthesia assist marketers and content creators?
These tools help with content creation tasks like writing, editing, and generating images and videos.
178
How do AI chatbots like Zendesk and LivePerson enhance customer service?
AI chatbots provide automated customer support, answering questions and resolving issues quickly.
179
What are Tableau and Power BI used for?
Tableau and Power BI are data visualization tools that help with analyzing and presenting data.
180
How does GitHub Copilot assist developers?
GitHub Copilot is an AI-powered code completion tool that helps developers write code faster and more accurately.
181
Name three AI tools that can help with task management.
Todoist, Microsoft To Do, and Evernote
182
What are ChatGPT and Gemini known for?
ChatGPT and Gemini are AI language models that provide versatile applications, including text generation, translation, and coding assistance.
183
What are the four phases of the Amazon AI Services Framework?
Data preparation
Model development
Deployment
Optimization
184
In the Amazon AI Services Framework, what does the data preparation phase involve, and what tools can be used?
It involves gathering, storing, and preparing data for AI applications. Tools include Amazon S3, AWS Glue, and Amazon Redshift.
185
Give an example of how an e-commerce company might use the model development phase of the Amazon AI Services Framework.
They could use Amazon SageMaker to create a recommendation engine and refine its accuracy through testing on historical data.
186
What is the focus of the deployment phase in the Amazon AI Services Framework, and what tools are helpful?
Integrating AI models into production environments. Tools include Amazon SageMaker, AWS Lambda, and Amazon CloudWatch.
Card 5
187
How does the optimization phase in the Amazon AI Services Framework contribute to continuous improvement?
By analyzing performance data, optimizing models for efficiency and accuracy, and scaling AI applications organization-wide using tools like Amazon SageMaker Debugger, Amazon Personalize, and AWS Step Functions.
188
What are the four phases of the OpenAI Framework?
Data Preparation
Model Development
Model Deployment
Continuous Improvement
189
In the OpenAI Framework, how is data prepared for AI applications?
Data is gathered, cleaned, and prepared using tools like the OpenAI API, Pandas, and NumPy, ensuring quality and governance.
190
Give an example of how a content company might use the OpenAI Framework for model development.
They could employ GPT-3 for generating marketing copy and refine it using Jupyter Notebooks.
191
What tools and technologies are used for model deployment in the OpenAI Framework?
The OpenAI API, Docker, and Kubernetes are used for seamless integration and scaling of AI models.
192
How does the continuous improvement phase in the OpenAI Framework ensure ongoing optimization?
By using tools like the OpenAI API, TensorBoard, and Google Analytics to gather feedback and optimize models based on real-time data.
193
What are the key areas where Facebook leverages AI in its platform?
Data processing, content recommendation, and user interaction.
194
How does Facebook ensure data privacy in its AI integration framework?
By adhering to data privacy regulations like GDPR and implementing secure data integration and processing practices.
195
What tools and technologies does Facebook utilize for AI model development?
Machine learning algorithms from FAIR's AI Research Tools and PyTorch are used for model training and validation.
196
How does Facebook deploy AI models for real-time applications?
By integrating them into news feed algorithms and ad targeting systems, using tools like the Facebook Developer API and AI Infrastructure for automated deployment and updates.
197
What is the purpose of continuous improvement in Facebook's AI integration framework?
To continuously optimize AI models and algorithms for better performance by analyzing user feedback and performance metrics, using tools like Facebook Analytics and AI Research Tools.
198
What are the key components of the data preparation phase in AI adoption frameworks?
Collecting, cleaning, and organizing data while ensuring data quality and relevance for model training.
199
What does the model development phase in AI adoption frameworks involve?
Designing and training AI models using algorithms and frameworks to create predictive models.
200
What is the focus of the deployment phase in AI adoption frameworks?
Integrating models into production environments and ensuring they are scalable and reliable.
201
How is the optimization phase crucial for continuous improvement in AI adoption frameworks?
By continuously monitoring and improving model performance and updating models based on new data and feedback.
202
AI algorithms
Programming that tells the computer how to learn to operate on its own, thereby enabling it to attain artificial intelligence (AI).
203
AI-driven predictive maintenance systems
Using AI in production control, to monitor and detect missing materials or quality issues.
204
AI-driven robotics
Robots powered with AI that are augmented with a variety of sensors.
205
AI-driven energy management systems
Using AI in production control to manage and optimize energy production, distribution, and consumption within energy systems.
206
AI-powered quality control systems
Using AI in production control, to manage quality by enhancing efficiency, accuracy, and decision-making capabilities in various fields.
207
Augmented intelligence
Technology that enhances human capabilities by providing AI-powered tools and assistance.
208
Authorization
The process of determining whether an authenticated user has the right to perform an operation.
209
Automation
The techniques involving technology, programs, robotics, or processes to achieve minimal human intervention.
210
Big data stores
A larger, more complex data set, especially from new data sources.
211
Big data
A dynamic, large, and disparate volume of data being created by people, tools, and machines.
212
Biometric
Application of statistical analysis to biological data of individuals by means of unique physical characteristics.
213
Carbon footprints
The greenhouse gases produced by digital technology resources, devices, tools, and platforms.
214
Chatbots
Computer programs that simulate human conversation to provide information, answer questions, and perform tasks via text or voice interactions.
215
Claude models
Advanced AI language models, designed for natural language understanding and generation. They are similar to OpenAI's GPT series and are used in various applications such as chatbots, content creation, and other AI-driven tasks that require language processing capabilities.
216
Cloud computing
A technology that stores and uses data and services over the internet instead of keeping everything on your computer.
217
Cobots or collaborative robots
AI-driven robots that can work alongside humans in a shared workspace to enhance efficiency and increase productivity. Unlike traditional robots, cobots are designed to interact directly with people, making automation more accessible and versatile.
218
Cognitive
The term involves intellectual activities such as thinking, reasoning, and problem-solving.
219
Cognitive computing
A technology that can evaluate an individual's performance and offer personalized recommendations. It provides enhanced functionality, adaptability, and intelligence across various domains.
220
Cyberattack
An attempt targeting to damage a computer network, computer information system, or personal digital devices.
221
Cybersecurity
Protection of information technology of individuals and organizations from cyberattacks.
222
Data analysis
Process that involves cleaning, transforming, and modeling data to uncover useful information to aid business decision-making.
223
Data augmentation
The process of generating data from the existing data to train machine learning models.
224
Dashboards
Tool to present a bird's-eye view of the complete picture while also allowing you to drill down into the next level of information for each parameter. These are easy to comprehend for an average user, make collaboration easy between teams, and allow you to generate reports on the go.
225
Dall-E model
A multimodal model developed by OpenAI that exhibits the ability to generate images that precisely match the input text or prompt.
226
Data scientist
Data professionals who develop algorithms, build predictive models, and uncover patterns and trends in large data sets. They apply statistical analysis, especially inferential statistics, machine learning, and predictive modeling, to extract insights from data and make predictions.
227
Data readiness
The technique involves removing inconsistencies, filling gaps, and ensuring the data is relevant to the problem of the data.
228
Deepfake
Altering of an image or recording to misrepresent someone as doing or saying something that was not actually done or said.
229
Deep learning
Deep learning is a specialized subset of machine learning. Deep learning layers algorithms to create a neural network, which is an artificial replication of the brain's structure and functionality.
230
Digital landscape
The digital landscape encompasses all hardware, software, and content involved in digital advertising, which is where businesses and customers interact.
231
Digital data source
The digital location where the data is held in a data table, object, or other storage format.
232
E-commerce
A platform for buying and selling products and services, transmitting funds and data over the internet.
233
Edge computing
The practice of processing data closer to the source of generation rather than relying on centralized data centers.
234
Edge AI
A type of AI that lives on the device itself rather than relying on the cloud.
235
Electronic health records (EHRs)
A digital version of patient-centered records that makes information available in real time.
236
Encryption
A method of encoding data so that only authorized individuals can comprehend the information.
237
Generative AI
AI that is capable of creating new content (text, images, music, audio, and video) and responding to human conversations.
238
Generative AI: Variational autoencoders (VAEs)
VAEs are a type of generative AI model that works by transforming input data through encoding and decoding. They have three main parts: an encoder network, a latent space, and a decoder network.
239
Generative AI: Generative adversarial networks (GANs)
GANs involve two neural networks: the generator and the discriminator. The generator creates new data samples, and the discriminator checks if the data is real or fake.
240
Generative AI: Autoregressive models
Autoregressive models create data sequentially, considering the context of earlier generated elements. These models predict the next element in the sequence based on the previous one.
241
Generative AI: Transformers
Transformers are generally used in natural language processing (NLP) tasks. They consist of encoder and decoder layers, enabling the model to effectively generate text sequences or perform cross-language translations.
242
Generative AI: Unimodal models
Unimodal models process inputs and generate outputs within the same modality.
243
Generative AI: Multimodal models
Multimodal models handle inputs from one modality and produce outputs in a different modality.
244
Hallucinations
A phenomenon where AI can produce outputs that are inaccurate or completely fabricated. The phenomenon highlights the importance of critical thinking when using AI-generated content.
245
Internet of Things (IoT)
A network of physical devices connected to the internet that collect and share data for processing and analysis. These devices can be sensors, cameras, or other devices that generate data.
246
Inventory management
The process of tracking, controlling, and usage of inventory from purchase to the sale of the goods.
247
Interactive learning environment
A dynamic learning setup where learners actively engage with content, instructors, and peers through digital or physical interactions to enhance learning.
248
Large language model (LLM)
A type of AI program that can recognize and generate text. LLMs are pretrained on large amounts of data.
249
Machine learning
The branch of AI and computer science focuses on using data and algorithms to imitate how humans learn, gradually improving accuracy. It helps systems learn and improve at forecasting, much like how people learn from experience.
250
Machine learning models
Computer programs that aim to train the computers to identify patterns within new data and make predictions.
251
Machine learning engineers
Programmers who construct the algorithms, systems, models, and frameworks that enable machines to learn and perform functions independently and effectively.
252
Meta's Llama models
An accessible, open-source LLM designed for developers, researchers, and businesses to build, experiment, and responsibly scale their generative AI ideas.
253
Natural language processing (NLP)
Machine learning technology that enables computers to interpret and comprehend human language.
254
Neural networks
Computational models that are influenced by the human brain's neural structure.
255
Neurons in AI
An artificial neural network consists of interconnected nodes known as neurons. The neurons take incoming data, like the human brain's neurological network, and learn to make decisions over time.
256
Neural network: Perceptron
The simplest type of artificial neural network consisting of only input and output layers.
257
Neural network: Feed-forward
A type of artificial neural network in which information flows in one direction, that is, in the forward direction. Each neuron in a layer receives input from neurons in the previous layer and then passes its output to neurons in the next layer.
258
Neural network: Deep feed-forward
A type of neural network that is similar to the feed-forward network with just more than one hidden layer.
259
Neural network: Modular
A type of neural network that combines two or more neural networks to arrive at the output.
260
Neural network: Convolutional neural network (CNN)
A type of neural network that is particularly well-suited for analyzing visual data.
261
Neural network: Recurrent neural networks (RNNs)
A type of neural network where the neurons in hidden layers receive an input with a specific delay in time. This allows the RNN to consider the context of the input.
262
IBM watsonx Assistant
A conversational AI solution that empowers a broader audience, including non-technical business users, to effortlessly build generative AI Assistants.
263
OpenAI
A private AI research laboratory that aims to develop and direct AI in ways that produce services.
264
Predictive analytics
A method that predicts future outcomes by using historical data combined with statistical modeling, data mining techniques, and machine learning.
265
Rehabilitation robots
Robots that help patients recover mobility and strength, offering personalized therapy.
266
Reinforcement learning
A category of machine learning technology that trains software to make decisions to achieve the most optimal results.
267
Robo-advisors
An online application that uses AI algorithms to offer automated, algorithm-driven investment suggestions, portfolio management, and financial planning services.
268
Robots
Machines with complex systems made up of several key components such as sensors, actuators, and controllers.
269
Robotics
A technology that involves designing, constructing, and operating robots, where the machines can perform tasks by themselves or with some help.
270
Robotic process automation (RPA)
A type of computer software that helps create, use, and control virtual robots. These robots act like humans when they work with digital systems and software.
271
Salesforce Marketing Cloud
An AI-driven marketing automation platform that uses machine learning algorithms to segment customers, target them with personalized messages, and optimize campaign performance.
272
Sentiment analysis
Processing digital text using AI to analyze and determine if the emotional tone of the message is positive, negative, or neutral.
273
Smart grids
An electricity network that uses digital technologies to monitor and manage the transport of electricity from sources to meet the varying electricity demands of end users.
274
Smart home
A technology that allows homeowners to control appliances, thermostats, lights, and other devices remotely using a smartphone or tablet through an internet connection.
275
Speech-to-text (STT)
A technology that changes spoken words into written text using neural networks.
276
Supervised learning
A category of machine learning technology that uses labeled datasets to train algorithms to predict outcomes and recognize patterns.
277
Strong AI or generalized AI
A type of AI system that replicates advanced human functions, such as reasoning, planning, and problem-solving.
278
Super AI or conscious AI
A type of AI system that is superintelligent by being self-aware and intelligent enough to surpass the cognitive abilities of humans.
279
Statistical analysis
The process of collecting large volumes of data and then using statistics and other data analysis techniques to identify trends, patterns, and insights.
280
Statistical methods
Methods useful to ensure that data is interpreted correctly and apparent relationships are meaningful and not just happening by chance.
281
Text-to-speech (TTS)
A type of assistive, "read aloud," technology that reads digital text aloud.
282
Unsupervised learning
A category of machine learning technology that uses algorithms to analyze and cluster unlabeled data sets.
283
Virtual assistance
AI-powered software applications designed to assist users with various tasks, such as setting reminders, providing information, and managing schedules.
284
Voice assistance
They can be interacted with using voice commands. Users can speak to these assistants to perform tasks, ask questions, and control other devices.
285
Video analytics
Use of advanced AI and machine learning algorithms to monitor, analyze, and manage large volumes of video.
286
Weak AI or narrow AI
AI that is focused on one narrow task.
287
What are LLM hallucinations?
Outputs of LLMs that deviate from facts or contextual logic. They can range from minor inconsistencies to completely fabricated or contradictory statements.
288
What are some types of LLM hallucinations?
Sentence Contradiction: An LLM generates a sentence that contradicts one of its previous sentences.
Prompt Contradiction: The generated sentence contradicts the given prompt.
Factual Contradiction: The LLM generates incorrect facts.
Nonsensical/Irrelevant Information: The LLM includes information that doesn't make sense in the context.
289
What are some causes of LLM hallucinations?
Data Quality: The training data may contain errors, biases, or inconsistencies. It may also be incomplete or lack coverage of certain topics.
Generation Method: The methods used to generate text (e.g., beam search, sampling) can introduce biases and trade-offs between fluency, diversity, coherence, creativity, accuracy, and novelty.
Input Context: Unclear, inconsistent, or contradictory prompts can confuse or mislead the model.
290
How can we reduce LLM hallucinations?
Provide clear and specific prompts. The more detailed the prompt, the better the LLM can understand the user's expectations.
Employ active mitigation strategies. Use LLM settings (e.g., temperature parameter) to control the generation process.
Use multi-shot prompting. Provide multiple examples of the desired output to help the LLM recognize patterns and context.
291
