Explainable AI
1 like12,089 views
The document discusses the rapid advancements in explainable AI (XAI), emphasizing the need for transparency and accountability in artificial intelligence systems to foster trust in their decision-making processes. It outlines various methods for providing explainable AI, including inherently interpretable machine learning approaches and new techniques for complex neural networks. Additionally, the document highlights the challenges and goals of explainability, such as enhancing model interpretability and fulfilling regulatory requirements.
Explainable AI
- 2. Factors driving rapid advancement of AI Third Wave of AI SymbolicAI Logic rules represent knowledge No learning capability and poor handling of uncertainty StatisticalAI Statistical models for specific domains training on big data. No contextual capability and minimal explainability. Explainable AI Systems construct explanatory models. Systems learn and reason with new tests and situations. GPUs, On-chip Neural Network Data Availability Cloud Infrastructure New Algorithms 2
- 3. There are two ways to provide explainable AI: • Use Machine learning approaches that are inherently explainable such as decision trees, knowledge graphs and similarity models. • Develop new approaches to explain complicated neural networks. What is XAI? It is a AI system that explains their decision making which is referred as Explainable AI or XAI. The goal of XAI is to provide verifiable explanations of how machine learning systems makes decisions and let humans to be in the loop. 3
- 4. What is Explainable AI? Black BoxAI Data Black-Box AI AI product Explainable AI Explainable AI Explainable AI Product Explanation Decision Feedback Decision, Recommendation Clear & transparent Predictions I understand why I understand why not I know why you succeed or fail I understand, so I trust you Confusion with today’s AI Black Box Why did you do that? Why did you not do that? When do you succeed or fail? How do I correct an error? 4 Data
- 5. Black-box AI Creates Confusion and Doubt Can I trust our AI decisions? How do I answer this customer complaint? Is this the best model that can be built? How doI monitor and debug thismodel? Why I am getting this decision? How can I get a better decision? Poor Decision Black-box AI </> </> Are these AI system decisions fair? Internal Audit, Regulators Data Scientists IT & Operations Business Owner 5 Customer Support
- 6. Why Do We Need It? 6 • Artificial Intelligence are increasingly implemented in our everyday lives to assist humans in making decisions. • These trivial decisions can vary from a lifestyle choices to more complex decisions such as loan approvals, investments, court decisions and selection of job candidates. • Many AI algorithms are a Blackbox that is not transparent. This leads to trustability concerns. In order to trust these systems, humans want accountability and explanation.
- 7. Why Do We Need It? 7 • While the machine learning systems deployed in 2008 were mostly within the products of tech-first companies (i.e. Google, YouTube), the false prediction would result in the wrong recommendation to the application user. • But, when it is being deployed in other industries such as military,healthcare, finance, it would lead to adverse consequences affecting many lives. • Thus, we create AI systems that explain their decision making.
- 8. 8 • Why did you do that? • Why not something else? • When do you succeed? • When do you fail? • When can I trust you? • How do I correct an error? We are entering a new age of AI applications. Machine learning is the core technology. Machine learning models are opaque, non-intuitive, and difficult for people to understand. AI System DoD and non-DoD Application Transportation Security Medicine Finance Legal Military User
- 9. Process of XAI AI is Interpretability. It collaborates between• The significant enabler of explainable human and artificial intelligence. • Interpretability is a degree to which a human can understand the cause of a decision. • It strengthens trust and transparency,explains decisions, fulfil regulatory requirements, and improve models. • The stages of AI explainability is categorized into pre-modelling, explainable modelling and post-modelling. They focus on explainability at the dataset stage and during model development. 9
- 10. Explainable AI Feedback Loop Train Deploy Monitor A/B Test Predict “Explainability By Design" For AI Products Model Diagnostics Root Cause Analytics Debug Performance Monitoring Fairness Monitoring Model Comparison Cohort Analysis Model Debugging Model Visualization Model Evaluation Compliance Testing QA Model Launch Signoff Model Release Mgmt Explainable Decisions API Support 10
- 11. Explainability Approaches 11 • The popular Local Interpretable Model-agnostic Explanations (LIME) approach provides explanation for an instance prediction of a model in terms of input features, the explanation family, etc. • Post-hoc Explainability approach of AI Model creates. • Individual prediction explanations with input features, influential concepts, local decision rules. • The global prediction explanations with partial dependence plots, global feature importance, global decision rules. • The build an interpretability model approach creates • Logistic regressions, decision trees, generalized additive models(GAMs).
- 12. Why Explainability: Improve ML Model Standard ML Interpretable ML Data ML model Model/data Improvement Data ML model Predictions Generalization error Generalization error + Human experience Verified predictions Interpretability Humaninspection 12
- 13. Explanation Targets 13 • The target specifies the object of an explainability method which varies in type, scope, and complexity. • The type of explanation target is often determined according to the role-specific goals of end users. • There are two types of targets: inside vs outside, which can also be referred as mechanistic vs Functional. • AI experts require a mechanistic explanation of some component inside a model to understand how layers of a deep network respond to input data in order to debug or validate the model.
- 14. Explanation Targets 14 • In contrast, non-experts often require a functional explanation to understand how some output outside a model is produced. • In addition, targets can vary in their scope. The outside-type targets are typically some form of model prediction. They can be either local or global explanations. • The inside-type targets also vary depending on the architecture of the underlying model. • They can either be a single neuron, or layers in a neural network.
- 15. Explanation Drivers 15 • The most common type of drivers are input features to an AI model. • Explaining an image classifier predictions in terms of individual input pixels can result in explanations that are too noisy, too expensive to compute, and more importantly, difficult to interpret. • Alternatively, we can rely on a more interpretable representation of input features knownas super-pixels in the case of image classifier prediction. • All factors that have an impact on the development of an AI model can be termed as explanation drivers. Explainer (LIME)
- 16. Explanation Families 16 • A post-hoc explanation aims at communicating some information about how a target is caused by drivers for a given AI model. • An explanation family must be chosen such that its information content is easily interpretable by the user. • Importance Scores - The individual importance scores are meant to communicate the relative contribution made by each explanation driver to a given target. • Decision Rules - Decision trees is where outcome represents prediction of an AI model and condition is a simple function defined over input features.
- 17. Explanation Families 17 • Decision Trees - Unlike decision rules, they are structured as a graph where internal nodes represent conditional tests on input features and leaf nodes represent model out- comes. In a decision tree each input example can satisfy only one path from the root node to a leaf node. • Dependency Plots - They aim at communicating how a target’s value varies as a given explanation drivers’ value varies, in other words, how a target’s value depends on a driver’s value.
- 18. • To explain pre-developed AI models, multiple methods have been proposed. • They vary in terms of their Explanation target, Explanation drivers, Explanation family and Extraction mechanism. • XAI is an active research area with new, improved methods being developed consistently. • Such diversity of choices can make it challenging of XAI experts to adopt the most suitable approach for a given application. • This challenge is addressed by presenting a snapshot of the most notable post- modelling explainability methods. Conclusion 18
- 19. To assist you with our services, please reach us at hello@mitosistech.com www.mitosistech.com IND: +91-78240 35173 US: +1-(415) 251-2064