AI

Question 1

What are Foundation Models?

Answer 1

Foundation models are large-scale pre-trained neural network architectures, like BERT or GPT, serving as bases for various AI tasks. They’re fine-tuned for specific applications like language understanding, generation, and more.

Question 2

Watsonx.ai

Answer 2

With watsonx.ai, you can train, validate, tune and deploy foundation and machine learning models with ease.

Question 3

What are the 3 components of the WatsonX platform?

Answer 3

Watsonx features watsonx.ai for foundation models and generative AI, watsonx.data for a flexible data store, and watsonx.governance for responsible, transparent AI workflows.

Question 4

Governed data and AI

Answer 4

It refers to the technology, tools, and processes that monitor and maintain the trustworthiness of data and AI solutions.

Companies must be able to direct and monitor their AI to ensure it is working as intended and in compliance with regulations.

Question 5

5 AI pillars of trust (trustworthiness)

Answer 5

• Transparency
• Explainability
• Fairness
• Robustness
• Privacy.

Question 6

Privacy

Answer 6

AI must ensure privacy at every turn, not only of raw data, but of the insights gained from that data. Data belongs to its human creators and AI must ensure privacy with the highest integrity.

Question 7

Robutness

Answer 7

An AI solution must be robust enough to handle exceptional conditions effectively and to minimize security risk. AI must be able to withstand attacks and maintain its integrity while under attack.

Question 8

Fairness

Answer 8

An AI solution means the reduction of human bias and the equitable treatment of individuals and of groups of individuals.

Question 9

Explainability

Answer 9

Simple and straightforward explanations are needed for how AI is used. People are entitled to understand how AI arrived at a conclusion, especially when those conclusions impact decisions about their employability, their credit worthiness, or their potential.

Question 10

Transparency

Answer 10

The best way to promote transparency is through disclosure. It allows the AI technology to be easily inspected and means that the algorithms used in AI solutions are not hidden or unable to be looked at more closely.

Question 11

Open and Diverse Ecosystem

Answer 11

The teams building AI solutions must be made up of people from different backgrounds and closely resemble the gender, racial, and cultural diversity of the societies which those solutions serve.

A culture of diversity, inclusion, and shared responsibility, reinforced in an open ecosystem, is imperative for building and managing AI.

Question 12

3 Principles of AI in an organization - Foundational Components of Ethics

Answer 12

• The purpose of AI is to augment human intelligence (not replace it)
• Data and the insights belong to their creator
• Technology must be transparent and explainable AI

Question 13

AI governance

Answer 13

It involves managing and overseeing AI processes, people, and systems to ensure they align with organizational goals, stakeholder expectations, and regulatory compliance throughout the AI lifecycle.

Question 14

Affinity bias

Answer 14

Seeking out or preferring those who seem similar to you

Question 15

Availability bias

Answer 15

Overestimating the importance of an event with greater “availability” in memory, like an event that happened most recently or was highly unusual

Recent exposure, emotions, or media influence make vivid data seem more significant, affecting decisions

Question 16

Confirmation bias

Answer 16

Seeking only information that confirms what you already believe is true

Question 17

Halo effect

Answer 17

Interpreting another person’s actions through a positive lens because you feel favorably toward them

Question 18

Status quo bias

Answer 18

Opting to maintain the current situation even when better alternatives exist

Question 19

Biases in AI

Answer 19

It’s a systematic error that has been designed, intentionally or not, in a way that may generate unfair outcomes. Bias can be present both in an AI system’s algorithm and in the data used to train and test the system.

Question 20

The 5 phases of the AI lifecycle

Answer 20

• Scope and plan
• Collect and organize
• Build and train
• Validate and deploy
• Monitor and manage

Question 21

Scope and plan

Answer 21

During the scope and plan stage, you define the project and evaluate potential ethical issues, including bias, by answering questions about topics like business expectations for fairness and transparency, regulation, and sensitive data handling.

Unconscious biases

Question 22

Collect and organize

Answer 22

During the collect and organize stage, you gather and prepare the data that will be used to train your model. Because models depend on data to learn, data quality is critical to AI fairness – yet there are many ways for bias to arise in this stage.

• Sampling bias
• Exclusion bias

Question 23

Build and train

Answer 23

During the build and train stage, you develop models and begin feeding them the data gathered and prepared in the collect and organize stage.

• Observer bias/Confirmation bias
• Aggregation bias

Question 24

Validate and deploy

Answer 24

During the validate and deploy stage, you evaluate the model’s performance and deploy it into production if it passes validation.

• Evaluation bias
• Deployment bias

Question 25

Monitor and Manage

Answer 25

During the monitor and manage stage, you monitor and manage the model after it is deployed for quality, performance, bias and fairness, and accuracy and data drift. ## Footnote - Evaluation bias - Deployment bias

Question 26

This bias: If you label or classify data according to your subjective judgment rather than objective guidelines, you risk imprinting your own perspectives on the system.

Answer 26

Confirmation bias/Observer bias ## Footnote if you labels images of lions as “cats” but your teammate does not, the system will be less accurate because of the inconsistently applied labels.

Question 27

This bias: Occurs when the data sample used to train a model is not representative of a population, overrepresenting or underrepresenting one or more classes in that population. Sampling bias can lead to inaccurate and sometimes discriminatory outcomes.

Answer 27

Sampling bias ## Footnote darker faces are often underrepresented in data used to train facial recognition systems.

Question 28

This bias: Occurs during data preprocessing, when a human removes seemingly extraneous or irrelevant data from the sample, skewing the data that trains the model.

Answer 28

Exclusion bias ## Footnote you might remove customer location data believing it to be irrelevant to a model, when in reality that location data might help the model detect geographically-influenced patterns.

Question 29

This bias: It happens when different populations are inappropriately aggregated in order to build a single model, even though that model might not fit all populations equitably.

Answer 29

Aggregation bias ## Footnote can often arise in medicine, when predictive models are built that ignore varying prevalence rates among different ethnic populations.

Question 30

This bias: Can occur after a model is built but before it is deployed, if it is assessed against benchmarks that are not representative of the population it is intended to assess.

Answer 30

Evaluation bias ## Footnote a model that predicts election outcomes might exhibit bias if it was evaluated using local election data but is deployed to predict national elections.

Question 31

This bias: If a system is used in a new or unanticipated context, it could generate biased outcomes.

Answer 31

Deployment bias ## Footnote a vision recognition system that is trained to recognize cars in the US might be biased when used in Europe, since the makes and models of American and European cars are often different.

Question 32

Benefits of Governance in AI

Answer 32

- Trust: Aligning AI with values builds transparent, fair, and trustworthy systems, enhancing client satisfaction and brand reputation. - Efficiency: Standardized and optimized AI activities accelerate development, reducing time-to-market. - Compliance: Managed and monitored AI activities ease alignment with industry regulations and legal requirements, reducing compliance burdens.

Question 33

Foundation model

Answer 33

An AI model that can be adapted to a wide range of downstream tasks. Foundation models are typically large-scale (e.g. billions of parameters) generative models trained on unlabeled data using self-supervision.

Question 34

Foundation model-based AI system

Answer 34

An AI system built on one or more Foundation Models.

Question 35

Generative AI

Answer 35

A class of Machine Learning techniques whose purpose is to generate content or data of many kinds such as audio, code, images, text, simulations, 3D objects, and videos. Foundation models are generative in nature.

Question 36

Ethics by Design (EbD) | IBM

Answer 36

A framework of principles and best practices designed to support the development of trustworthy AI and other technologies at IBM

Question 37

Goal of Ethics by Design | IBM

Answer 37

Support IBM's commitment to responsible development and deployment of technologies in all IBM products and services across all business units and geographies.

Question 38

AI Ethics Board

Answer 38

It's a committee of diverse and knowledgeable stakeholders from across the organization whose responsibilities is to enforce Trust and Transparency Principles, prioritizing awareness, education, and accountability for ethical AI development and deployment within the company's framework and values.

Question 39

What is AI Ethics?

Answer 39

A multidisciplinary field that studies how to optimize AI's beneficial impact while reducing risks and adverse outcomes.

Question 40

AI Lifecycle

Answer 40

The AI lifecycle consists of a sequence of stages that bring together various personas and best practices, which include: - Scope and plan - Collect and organize - Build and test - Validate and deploy - Monitor and manage