Skript Flashcards
(97 cards)
1
Q
How do we define “artificial intelligence”?- Def.
A
Intelligence exhibited by non-biological systems
2
Q
Artificial General Intelligence (AGI))- Def.
A
Human-level AI. The computer would need to be able to integrate many capabilites
3
Q
Narrow AI-Def.
A
Ability to accomplish a narrow & well-defined set of goals (Matching resume to open job positions)
4
Q
Cognition-Def.
A
the mental action or process of acquiring knowledge and understanding through thoughts, experience, and the senses. The psychological result of perception and learning and reasoning
5
Q
Neural Networks -Def.
A
trained using machine learning approaches form the basis of most attempts at conginitc computing
6
Q
Neuromorphic hardware -Def
A
It is being worked on in parallel to purely software.
The largest artificial neutral netwworks may have ca. 200 layern=> size of a frog brain
7
Q
Artificial intelligence => ML=> DL
A
[AI] Human intelligence exhibited by machines - terms for (narrow) use cases, includes any simulation of human intelligence => [ML] Computer can learn from data without being explicitly programmes - terms for the technology, Uses numerical and statistical approaches (pattern), Models are built using “training” computation runs, can also train through usage => [DL aka Neural Network] A subfield of ML that uses specialized computational techniques, typically multi-layer (2+) artifical neural network, layering allows cascaded learning and abstraction levels (e.g. line recognition -> shape -> object -> scene ), Computationally intensive enabled by clouds, GPU’s, and increasingly more specialized HW
8
Q
AI is becoming feasible because of a conflux of several technological advancements. The field is advancing rapidly beyond mathematical proofs because researchers can experiment with data and methods:
A
Powerful hardware -GPUS/TPUs accelerated training time of DNNs
Digitization & Big data- e.g. IOT, Digital receipts, Technologies like SAP HANA
Cloud computing -access powerful hardware with an Internet connection and credit card
Advances in applied deep neutral networks- Deep learning, new statistical techniques
9
Q
What are First AI Attempts?
A
Rules-based “expert systems”:
If-then-else rules or algorithms, highly logical, structures & explainable, use structured, codified data
10
Q
What is probabilistic machine learning
A
learn from data without being explicitly programmed (Cat-dog example)
11
Q
Summary and Key Takeaways
A
- Intelligence is the ability to achieve complex goals. AI is intelligence exhibited by non-biological systems
- Currently the most feasible approach to AI is machine learning -> learning from large amount of data without beeing programmed
- AI leverages many different statistical modeling approaches from simple regressions to deep learning networks. Many are years old. Rules-based expert systems are still a useful and valid approach
- Feasible today because of ..(you know)
- People are comfy with rules-based systems but less comfortable with probabilistic
- we all train machine learning models every day. Big advantage
12
Q
Basic ML Workflow
A
Trainings Data->Learning Algorithm -> Model1
Testing Data->Model2Accuracy Estimated
Model2->Model3 Prediction
13
Q
Machine Learning is bases on 3 different categories of approaches to learning:
A
Supervised learning:
We provice the machine with examples of the desired solutiom (most common)
Unsupervised Learning: Data is set unlabeled- machine finds pattern /strucutes, Evaluation is qualitative or indirect
Reinforcement learning: Intelligence agents learn a polivy given neither data nor labels, only rules in their environment, observes environment, performs actions, gets rewards/penalties, Not to be confused with retraining a supervides learning model
14
Q
Supervised learning:
A
In supervised the training inputs and desired outputs (called “labels”) are given by a “teacher”
15
Q
Unsupervised learning
A
No labels are given to the learning algorithm, leaving it on its own to find the structure (costumer segments)
16
Q
Reinforcement learning
A
we don’t start with data, but rather a defined envorinment. The agent trains its own policy throgh a process of trial and error
17
Q
Classes of problem AI can adress
A
- Predicition, Forecasting, PAttern detection
- Computer Vision & Sensory Information
- Natural Language Processing and Unstructed Text
18
Q
Regression (supervised) Def.
A
The output (i.e. what the machine predicts) is continours rather than discrete. Many kinds of regression models exist e.g. Year an employee will stay with the company
19
Q
Classification (supervised) Def.
A
The model assigns inputs into 2 or more discrete classes. Inputs are aaigned to call with a given confidence level-Most machine learning adresses classificaiton problems e.g. employee will churn or not?, customer will choose product A oder Product B (or C)
20
Q
Clustering Def. (mostly unsupervised) Def.
A
Divide the inputs into groups. The group are normally not know beforehand, making this typicaly an unsupervised task. E.g Customer Segmentation, Genome squencing
21
Q
Dimensionality Reduction Def.
A
Map from higher dimensinal features space to lower dimensions is uses to combine several highly similar features into a singel feature E.g combine several items on a survey into a single “construct” variable
22
Q
Examples for Models and Tools of statistical Model
A
Classification Analysis, Regression, Cluster analysis, Time series analysis, Association analysis, probability Distribution, Outlier Detection, Recommender Systems, Link prediction, Statistical Functions, Data prepation
23
Q
How do you do feature engineering?
A
Creation of categorical or dummy variables, data cleaning, feature normaliazation /scaling, handle missing values, standardize input format
24
Q
Misconception: A machine learns immediately after every click
A
Correction: A model is updated after we retrain model using the entire data set from scratch. Small exceptions include trransfer learning and active learning
25
Misconception: every action "reinfoces" and improves the model
Correction: reinforment learning is not training a supervised learning model. These are completely different approaches
26
Misconception: A model with the highest accuracy is the best model
Correction:Models with very high accuracy might be overfitted. When they are exposed to new data, te perfomance can degrade. Robustness and scability are equally imrportant.
27
Misconception: If we dont have training data, we can simulate data
If we know the functions and distribution already needed to simulate data, the we wouldnt need machine learning. Simulated data is viable in very rare circumstances (e.g. images created and roatetd using CAD software)
28
Misconception: Neural network are "superior" moddeling approaches
Neural Networks are appropriate (or the only possibility) for certain problem, but often "statistical" model are as accurate, more robust, and interpretable. Many models can be used to address the sam problem since performance depends on information in the data
29
Misconception: We require huge datasets to train a model
We don't need until we test it. Some model train well with less than 100 others require more nach 100.000s. depends if the data contain a signal /on data quality
30
Misconception: If it doesn't have 100% accuracy, we cannot use it
Nothing is 100% accurate" Unrealistiv expectation. AI resuts are probabilistic, exactly like results generated by humans (and AI is often fatser and more accurate..)
31
Misconception: We can explain all AI models
Many model types (esp, neural networks) are black boxes. Model interpretability is an active area of research in computer science - math problem rather than just engineering.
32
Misconception: We can generate actionable "insights" from the features in a machine learning model
AI interpretability (prediciton) is not explanatory modelling" Be careful.Two variables can be good predictors (correlation) but have the opposite or no relationsship in an explanatory model. Because X predicts Y is not to say if the change X it will Impact Y in the real world
33
Machine learning has several outcomes- which ones?
Regression, Classification, Clusterin and dimensionality reduction
34
What involves feature engineering
Cleaning, data, selecting relevant features, and codifying and tranforming the variables
35
Why are chatbots more feasible?
People are comfortable chatting with bots (messaging apps)
Performance of bots has increased substantially (largely thanks to neual networks /NLP)
Many benefits, Bots are: 24/7 aviable, efficient, cost reducing, multi-channel
36
What is the chief capability of conversational AI?
Assigning expressions (human sentences in unstructured text) to intens (categories of concepts that the bot understands)
37
Conversational AI - what are expressions ? What are intents?
Expressions: different sentences people say to communicate what they mean
Intents: Represent a concept that the bot will be able to understand by analyzing the user expressions, i.e. what the bot should know. Intents recognition is a core capability of a chatbot
38
What can NLP do?
NLP can recognize entities based on context and memory of the conversation. Simple to human but very difficult for computers. It extracts information from an expression (NER, sentiment analysis, language detection) and assigns it to an intent given confidence level. Responses are typically hard.-coded
39
Why do we build chatbots?
We build a chatbot and expose it to a website.
40
What are the 3 relevant components of a bot
The bot builder to create the conversation flows and skills, the NLP engine and the bot connector
41
Why do we typically hard-code responses of an AI Chatbot?
Natural Language generation (NLG) is an active area of AI research. I has some success with limited scenarios, but als carries the risk of of returnin nonsense (Like human customer servie agents, creating concise responses provides end user with a clear, consistent experience
42
What are examples of successful chatbot implementations?
Chatbots who do not answer simple FAQs. They can also pull data and execute transactions in back-end systems
43
Conversational AI and Hands-On exercise: Summary and Key Takeaways
1. Conversational AI functions mainly by natural language understanding and named entity recognition. Future systems my include natural language generation
2. Chatbos may be able to categorize sentiment and detect emotions, depending on training data and how advanced the underlying NLP is.
3. Chatbots are like people: they know as much as you teach them and require continuous governance training, and upskilling
44
Classes of problems AI can address
1. Prediction, Forecasting, Pattern Detection: Recommendations, Classification, Clustering, Anomaly/outlier detection, Forecastin, Matching, Proposals
2. Computer Vision & Sensory Information: OCR, Image classification, face detection, emotion recognition, signal processing (sounds, IoT data)
3. Natural Language Processing and Unstructured Text: Speech-to-tect, text-to-speech, Chatbots/ conversational AI, Sentiment analysis, Topic detection, Neural Machine translation, NER, NLG, Automated summerization
45
How does AI impact organizations
1. Automate reptitive tasks example: Chatbot answer website FAQs
2. Augment user experience example: Recommandation systems to auto-fill formd
3. Enable new business models example: Drones using computer vision to investigate buildings safety
46
Example of use cases for automation
1. Machine learning algorithms analyze content of unstructed ticket text from various channels
2. The service analyzes sentiment, classifies the ticket and routes to the service team
3. Experts and solutions are recommended to help he agent answer inquiries
47
What ML modellling capabilites requires a service ticket processing?
1. Classifier: Character level based convolutional neural Network (CNN)-categorization based on unstructured Text
2. Language Detection (classifier)
3. Sentiment analysis (classifier)
4. Named entity recognition, e.g. Date-time, location etc.
48
What could be the future of human-computer interfaces?
Using augmented reality, computer vision, conversational AI, and deep learning to scan an invoice and execute a transaction in an ERP system
49
What can computer vision do?
Aid the visual inspection of material on a production line. In reality we may combine several types of input, including photos but also data from infrared cameras, sensor data like weight, etc.
50
What do I need to know about Image-based material recommendation?
It is a common use case although it requires excellent training data
Image of material=> Image feature extraction similarity scoring>> Part identification and automatic classification
51
What is transfer learning
Refers to a method to reused a model developed for one task as the starting point for a model on a second, closely related task. It can accelerate training and reduce compute time
From generic inception model=> we exten with our logo picture => to improve perfomance
52
Summary and Key Takeaways: enterprise Application of AI
A. AI can (1) automate repetitive tasks, (2) transform how we interact with IT systems, and (3) create new business models
B. AI allows computers to work with unstructued data like image, sounds, and text, which in turn expands the reacht of enterprise IT into areas normally managed by humans
C. Large-scal automation of business processes is like "death by 1000 papercuts"-every step and use case
Exmaples of AI use cases in Business are numerous
53
What is strategy?
A coherent mix of policy and action designed to surmount a difficult, high-stakes challenge. Diagnosis=> Guiding Policy => Coherent Actions
54
What is digital transformation
Digital transformation is capturing events and artefacts as data, and using that data intelligently to change how things are done.
55
Why are companies deinvesting AI?
Because it is a hyped technology topic and expectations could meet reality
56
Implications for strategy management in your organization?
Dont put all AI into the same bucket, invest wisely in order to build and maintain capabilities, AI-based tranformation is like "deat by 1000 papercuts"- many small automation stepps will eventually combine to transform a company
57
How should companies entry the AI market?
Often companies run into roadblocks despite early successes with horizon 2 technologies. They need protection and attention to ensure they habe best market entry without being burnt due to lack of product quality. Word of mouth and references are key. Comapnies can only cross the chasm by targeting a specific niche first, where they establish a strong claim before expanding
58
What is the Innovator's Dilemma?
At first, emerging (disruptive) technologies are not profitable, perfomant, or scalable, Big companies cannot help themselves to keep investion in their cash cows, called sutaining technologies
....until eventually the emerging technology matures and knocks them out of the market
59
What decision does organizations face in AI business?
Make, ally or buy
60
How do you start building an AI?
Building and training an AI mdel is only part of the process (ca. 20% of effort). Most organizations struggle with identifying use cases, accessing data, and putting models into production with a sutainable governance (80% of the effort)
61
What is the typical AI project lifecycle?
1. Identify feasible use cases:
- identify feasble use cases (with ROI), find skilled data scientists and developers, access, explore, and clean the data, build, test models (specification, experimentation)
2. Put into production:-define architecture, integrate into a business appliction, create a data pipeline for training and production, monitor performance&scalability
3. Manage lifecycle &scalability:-create sharable features, repositories, documentation, and data pipilines, ensure model auditability, versioning, and repoducibility, Manage software, /SaaD lifecycle and landscape, retrain, decommission as needed
62
What are most AI use cases for?
Most AI use cases are for process optimization, while the main data source for AI in the DACH region is SAP systems, followed by production and machine data
63
Is machine learning appropriate for every challenge? What are the options?
Machine learning is not appropriate for every challenge. Many challenges are better addressed with rules, people/organization, or standard software developments. See figure on script 2.2 p.19)
64
What mean multidisciplinary in AI?
The required multidisciplinary makes AI a challenging field,
Predictive analytics is similar to machine learning. Both rely on similar statistical concepts and algorithms, ML is a bit broader in its application
ML requires capabilites in several areas: statistical modelling, computer programming, data manipulation + understanding of business context and use cases
65
How can the extend core enterprise applications?
Often we are ectending core enterprise applications with AI using a "bi-modal" IT approach, decoupling custom development from standard core processes. See figure in script 2.2 slide 21/22
66
Why do Ethics and AI exist?
Recent high-profile failures of AI in Leading Companies have raised serious questions about AI ethics: How bas impacts users, use of the technology, monetoring and controlling of the models, etc.
67
AI ethics and Governments
Governments are moving quickly to establish regulations concerning the societal, politval, and economic impact of AI. However policy makers often dont undertsnad AI, leading to regulatory uncertainty.
1. Regulations are still trying to figure out how to approach AI
2. Polify topics include ethical AI, investment (education, R&D funding), application of AI in the military and warfare, intellectual propery right, worker welfare and training
3. Unknown regulatory future is a potential risk, e.g. AI in validted business processes in the pharmaceutical industry, data access and protection
Therefore, organizations should establish an AI ethics policy and well-documented, scure, auditable governance process for AI use cases
68
What factor would politcians, executives, and economists most like to improve?
Ratio between volume of outputs and volume of inputs (work, capital). i.e. how many widgets can be produced given a unit of input.
Economic growth has two main drivers:
Population growth
Productivity growth: labor productivity, capital efficiency
- Labor productivity is stagnation
- returns to capital are increasing much fast than wages
-widening income inequality & labor market polarization
- who will buy products & Services if they have no money?
69
Why is labor productivity stagnation
The 1st (machanization), 2nd (mass production & electricity, and 3rd (IT) industrial revoluitons increased productivity, fueled growth, and improved standards of living. However, labor productivity and growth in developed nations is stagnation.
Solow's productivity paradoc: " you can see the computer age everywhere but in the productivity statistics!. Sometimes the ROI on technology investments is not easy to measure since retunrs can lag and depend on externalities to be realized
70
AI impacts work along 4 patterns of technological disruption. Can we thing of other examples?
```
Elimination (telegraph operators dont exist anymore)
Substituion/Redistribution (Truck drivers -> autonomous vehicles, Warehouse pickers -> Kiva robots (amazone)
Complementarity (Replacing paper-based quality assurance on the produxtion line with iPads, AI-&AR-Supported warehouse pickers)
Novel work (Data scientist, machine learning engineer, spaceship pilot, social media ambassador)
```
71
What should companies do with routine tasks?
Companies and policy makers should start identifying routine tasks and associated skills now.
Routine (physical or cognitive):
High risk of displcement
Low-value add, repetitive, mind-numbing
Non-routine (physical or cognitive):
More creativity, more knowledge, more flexibility and higher-order problem solving skills
different skills altogether
Empargy
constant re-training
genereally higher salaries, more value-add
72
What creative destruction were unleashed by technological innovation and automation?
Did cause job losses for some
did redistribute waelth, mostly to the machine and factory owners
But als freed people up to work on value-adding tasks
Lowered the production cost of items like knitted caps
Like an incoming tide raising all boats, increased the standard of living for everyone
73
What does the the 4th industrial revolution do for certain?
It disrupt work but also improves lives. The furute requires creativity, empathy, and constant reinvetion We are living in exciting times, so lets embrace the uncertainty and challenges in front of us
74
What guidelines are coming up for ethical AI?
Transparency: explaninibility why AI made a decision,
Justice, fairness, equality: avoiding bias and discrimination, right to appeal decisions, ensuring human rights, impact on the workforce etc.
Non-Maleficence: ensure safety and security, AI shoudl not harm people or introduce risk
Responsibility & accountibility: legal liability, actiong with integrity and foresight
Privacy: personal-identifiable data storage and usage (GDPR)
75
What does AI Transparency mean?
Means that stakeholders have a right to know whether AI is behind a decision, how it came to the decision, and what/how data is being used
76
What does AI Justice. fairness, equality mean?
include avoiding biased decision making in AI, having the right to appeal decisions, and avoiding an unjust impact of AI on broader society (e.g. distribution of waelth)
77
What does AI Non-Maleficence mean?
Means that AI should not cause harm or introduce unnecessary risks to security and well-being
78
What does AI Responsibility & accountability mean?
Takes into account legal liability, action with integrity, and ensuring that adequate processes are in place to reduce the risk of harm.
79
What does AI Privacy mean?
Considers if personally identifiable information (PII) is used in the AI applications and what inferences the model makes about an individual.
80
What are the keys to establish ethical AI in an organization?
Governance, technical standards, and best practises are key to establishing ethical AI in an organization
81
What does Governance for AI mean?
- AI ethics committee and AI ethics in steering committes sign off on use cases based on guidelines and a risk assessment
- Closesly monitor evolving policy and legislation regarding ethical AI
_ every AI use case in production should be transparent and auditable throughout its lifecycle: data collection and use, who trained the model, documented pipeline, perfoamnce metrics, continous monitoring
82
What does technical standards in AI mean?
Create a reference architecture for AI to ensure data security , centrailization, and avoidance of maverick projects, Build interpretable AI where possible, guidelines for data storage and use of vendors (e.g. APIs, Startups), and requirements for clear documentation
83
What does best practises for AI mean?
evaluate possibility of (selection) bias in your model and snrure GDPR compliance
consider what would happen to you organization if the model was leaked to the public or made incorrect predictions
84
AI ethics: Summary and key takeaways
1. AI ethics is an emerging topic with little consideration in most enterprises
2. Ethical AI policy discussions may lead to unforeseen regulations and compliance issues. Furthermore, plocy maker have difficulty understanding AI
3. AI is a core driver of the 4th industrial revolution. It will disrupt the workforce, redistribute wealth, and usher in a new era of increased productivity thanks to the automation of low-value tasks
4. Requirements of ethical AI center on transparency, justice and fairness, non-maleficence, responsibility, privaicy
5. Organizations are often ill-prepared to implement ethical AI requirements. Some issues, lie interpretable AI" and detection of bias, are emerging scientific areas without a simple solution:
6. AI ethics councils, creating an organizational position paper connecting values to AI ethics, careful, selection of use cases, screening training data are some measures that organization can put in place.
85
What are convolutional Neural Network
```
Features unknown, Features engineering = Cleaning, sizing images
Hyperparameter tuning (# of layers, epochs, loss function...)
```
86
What do you we do if we have highly unstructured data like photos?
Labelled data input -> Concolutional Neural Network -> Out (binary classification)
87
Definition Deep Learning
The focus in deep learning is on hyperparameter tuning
88
Statistical ML algorighms (procedure)
Training input -> processing -> feature engineering -> processing -> learnning -> model------ for smaller Data sets
89
Approach on deep learning:
Training input -> Processing -> architecture engineering (+some feature engineering) -> model -------for bigger data sets
90
Perfomance metrics df.
evaluate how well the model is fit against the trainind and validation data
91
Confusion matric def,
Shows correctly and falsely predicted categories e.g. performance metrics for a simple binary classification problem (folie 36, true false, ture positive)
92
What are entities?
Important keywords extracted from expressions. Entities are often a function of context and memory
93
Skills are made of three parts
Triggers
▪ Conditions that determine if the skill should be activated (= intents detected).
Requirements
▪ Determine the information that the bot needs to retrieve from the user, and how to retrieve it.
Requirements are normally data extracted as entities from a sentence.
Actions
▪ Performed by the bot when all requirements are complete (for example, send a message). This is hardcoded programming, e.g. respond with text, send data to a business application, etc.
94
Skills are made of three parts
Triggers
▪ Conditions that determine if the skill should be activated (= intents detected).
Requirements
▪ Determine the information that the bot needs to retrieve from the user, and how to retrieve it.
Requirements are normally data extracted as entities from a sentence.
Actions
▪ Performed by the bot when all requirements are complete (for example, send a message). This is hardcoded programming, e.g. respond with text, send data to a business application, etc.
95
Example use case
Input: Law document -> SAP MAchine learning: Language detection, Text classification, Text feature extraction, similarity scoring <=> Output: Alerts experts, suggests relevance, translates, highlights relevant text and learns continously
96
How do you develop strategy?
Diagnosis <=> Guiding Policy <=> Coherent Actions
97
Bi-model IT?
Decoupling custom development form standard core processes
