Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing
Download as PPTX, PDF0 likes19 views
This document discusses the integration of explainability-by-design for artificial intelligence (AI) in manufacturing, emphasizing the importance of transparency in AI decision-making processes. It outlines the challenges faced by complex machine learning models and presents the x-by-design approach as a solution to enhance understanding and trust in AI systems. The project is funded by the European Union's Horizon 2020 research and innovation programme, highlighting its significance in improving manufacturing efficiency and safety through explainable AI.
Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing
- 1. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 957362 Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing April 11th, 2024 – Chania, Greece A. Capaccioli, C. Agostinho, V. Antonello, F. Lampathaki and M. Sesana
- 2. • Artificial intelligence (AI) refers to systems that display intelligent behaviour by analysing their environment and taking actions – with some degree of autonomy – to achieve specific goals. • AI-based systems can be: • Software-based, acting in the virtual world (e.g. voice assistants, image analysis software, search engines, speech and face recognition systems) • Embedded in hardware devices (e.g. advanced robots, autonomous cars, drones or Internet of Things applications) • AI systems and Machine Learning (ML) have become ubiquitous in many scientific fields during the last decade • More data • More GPU power Artificial Intelligence Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Source: https://www.coditation.com/blog/decade-of-artificial-intelligence-a-summary
- 3. • Predictions show that this interest in AI systems will keep rising in the following years • Some major challenges: • The more difficult problems, the more complex AI gets, hence, less understandable by humans • In an industrial setting, such as manufacturing, ML models are often dependent on many variables and complex relations used to optimize processes, increase efficiency, reduce costs, or improve safety • Causal AI, AI-Trust, Human-centric AI, Explainable AI (XAI) is an emerging field • Provide insight into the underlying logic of models and improve transparency in the decision-making process Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Artificial Intelligence
- 4. Why is Explainable AI important? “Why did the AI system make this specific prediction or decision?” “Why didn’t the AI system make Prediction B for the same circumstances?” “What should I change in order for the AI system to decide something else?“ How to avoid undetected bias, mistakes, and miscomprehensions creeping into decision-making? How to ensure fair decision making without compromising security and privacy? How to facilitate robustness, accuracy and performance without creating additional liabilities? How to provide really actionable insights? I. Increase Human Trust in AI II. Increase Transparency and Reliability of AI If humans do not understand why/how a decision/prediction is reached, they shall not adopt/enforce it…
- 5. • Often in machine learning such property is not available, and to have it, models would need to be retrained which is a labour- and computation-intensive process • An explainable design requires that qualifiers can be extracted from data to describe heuristically meaningful features and assign causal relationships between inputs and outputs Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Explainability
- 6. Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Classic Explainable AI Models • Are those that are interpretable by themselves. Main properties: • Simulatability -> Human is capable of replicating • Decomposability -> Explain all model's parameters • Algorithmic transparency -> Understand how output is generated • Examples: • Linear, Generalized Linear and Additive Models • Decision trees • Rule-based Models • k-Nearest Neighbours (KNN) • Due to their simplicity, these models fail when dealing with complex problems such as the ones present in an industrial setting
- 7. • Methods used to explain the behaviour of complex ML models (black box) after they have been trained, bringing insights on how a model reaches its results • model-agnostic • model-specific Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Post-Hoc Explainability Techniques • Depending on the model, different explainability techniques can/should be experimented to get the best results • Model-agnostic techniques offer flexibility by separating the explanation technique from the ML model itself but may sacrifice efficiency or accuracy
- 8. • X-by-Design is the next paradigm shift and is a main exploitation result of the XMANAI European project • By prioritizing explainability from the outset, AI designers, engineers, and data scientists can construct models that inherently offer transparent insights into their decision-making processes, thereby mitigating the need for complex and often unreliable post hoc explanations • Support explainability in data, models and results Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing X-By-Design Paradigm
- 9. Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing X-By-Design Approach • Concrete understanding of data in terms of semantics and structure (data types) • Mapping to a common data model. • Interactive data exploration and visualization that can be leveraged to monitor potential data drifting or bias issues • Understanding the different AI (ML/DL) models towards global interpretability (answering how does a model work for all our predictions) and local interpretability (answering how a model is generating a specific prediction, given specific data points) • Use models natively easier to understand or apply different Post-hoc, such as explanations by simplification, feature relevance. Some common techniques include LIME, SHAP, etc. • Promotes shared understanding of results and translating them into concrete actions in an appropriate style/format • Fit-for-purpose visualizations act as the primer for communicating results to the involved stakeholders; Text explanations convey how to take action; Explanations by example support understanding; Counterfactual Explanations that aim to find the model’s decision boundaries
- 10. • XMANAI has also delivered appropriate assessment methods and techniques to address a number of complementary challenges that currently constitute significant data scientists’ pains • XAI Models Security that performs a risk and vulnerability assessment over different Explainable ML/DL models to offer immunity and robustness • XAI Models Performance, in terms of model metrics and explainability indicators • XAI Assets Sharing, to enable collaboration Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing X-By-Design Approach
- 11. • X-LEARN – consultancy services for defining/identifying the best methodologies or models to integrate explainability in the design phase of end-users’ projects • X-LEARN-basics, state-of-the-art research on the explainaiblity concept and on the identification of XAI models to be integrated in solutions to improve technological transparency and user experience • X-LEARN-applications: replication of project use cases to similar applications with other subjects, sharing the obtained results to strength user trust in AI. • X-APPLY – analysis of the AI status and most effective path to integrate digital solutions based on explainability in industrial contexts by providing IT services. X-By-Design Services Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing
- 12. Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Take-up Methodology
- 13. • XAI to help operators and engineers choose the best decision at any given moment (e.g. size, mix, and schedule of batches) so that better plant performance is achieved • XAI to forecast demand accurately and implement a Direct-to- Consumer strategy • XAI to quickly identify and understand production line or machine problems that affect operational capacity • XAI to increase the efficiency in the definition of a metrology measurement plan definition. Also supporting junior metrologists XMANAI Demonstrators Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing
- 14. • It is crucial to identify the benefits that a XAI system provides compared to a conventional AI system. • X-by-design approach involved user research activities to define main activities and to consider the contexts in which the application will be used, the definition of users' needs, and the associated explainability requirements, their prioritization in the development phase, and interfaces prototyping • Thanks to the use of interpretable models and transparent learning techniques, the X-by-Design approach provides detailed insight into the logic of AI, making easier to understand how it works and the parameters used, enabling companies to take full advantage of the benefits of AI in a transparent and understandable way, starting from system design. • The X-by-design paradigm in fact strongly relies on the human-machine interaction by bringing the focus on a human-friendly approach to explaining the model and making both manufacturer and customer interaction better. • Even though the path towards X-by-Design is under construction and presents its own challenges, the positive impact that can have on the overall manufacturing sector plays a crucial role from an industry perspective in boosting productivity and from a technological point of view in promoting XAI uptake. Integrating Explainability-by-Design for Transparent and Efficient AI in Manufacturing Challenges and Future Perspective
- 15. This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 957362 Thank you for your attention! carlos.agostinho@knowledgebiz.pt
Editor's Notes
- #4: AI trust, risk and security management (TRiSM)
- #7: Linear models compute the output as a linear combination of the input features following a formula. The computations for determining the output given the input are transparent and easily simulated by a human. However, in many cases, real world data cannot be modelled under these assumptions, hence generalised linear models [6], such as logistic regressions that allow non-continuous or unbounded output variables, or their extension to Generalised Additive Models (GAM) that are better-fit to model more complex input-output relationships Decision trees follow the idea that a model can be trained by splitting the data space in a recursive manner and then fitting a simple prediction model at each division obtained. Rule-based models identify a set of relational rules that can lead to a decision or prediction based on the data provided. If-then; fuzzy, support vector, Bayesian modelling aim to establish a probabilistic connection between the features and the output. Usually, Bayes models derive the posterior probability because of two antecedents: a prior probability and a likelihood function KNN models models rely on the metric to calculate the distance between the new sample and its neighbours
- #8: Model-agnostic techniques decouple the explanation technique from the ML model [19]. This approach has the great advantage of removing all restrictions when choosing which model will be used in a specific experiment. Furthermore, designing an accurate and efficient model agnostic method can be exploited by all black-box models. Sometimes they lack in efficiency or accuracy compared to the model-specific alternatives