Computer Sciences and Information Technology
Title : | Bayesian Optimization-based Neural-Symbolic Artificial Intelligence (BONSAI) for Explainable Learning from Urban Data |
Area of research : | Computer Sciences and Information Technology |
Focus area : | Artificial Intelligence, Smart Cities |
Principal Investigator : | Dr. Monidipa Das, Indian Institute Of Technology (Indian School Of Mines) Dhanbad, Jharkhand |
Timeline Start Year : | 2023 |
Timeline End Year : | 2025 |
Contact info : | monidipa@iitism.ac.in |
Equipments : | Server Computer |
Details
Executive Summary : | The increasing use of deep learning techniques for urban data analytics has led to the development of explainable artificial intelligence (AI) or XAI. However, deep learning models are often opaque and difficult to interpret, debug, and certify by non-technical individuals. This has led to the development of various models to address these issues. However, recent reviews on XAI models, particularly those in the context of urban computing, show several research gaps. Firstly, there is no concrete definition or structured format for explaining explainability on deep learning-based urban data analytics. Secondly, the explanations provided by existing models are limited in meaning and lack transparency and fidelity. Lastly, there is a lack of metrics to quantify the appropriateness of explanations for XAI goals. To address these issues, a research project aims to develop a unified framework that provides explanations covering all major dimensions, including justification, improvement, control, and discovery, along with metrics for attributing explanations and evaluating moral/ethical standards. The key component of this framework is a novel explainer built on Bayesian optimization-based neural-symbolic artificial intelligence, called BONSAI. BONSAI aims to develop a symbolic rule-base as a human interpretable representation of the complex DNN model, with Bayesian optimization optimizing the rules for better understandability. The inference engine applies this rule-base to explain how the model reaches a conclusion and generate new rules to enhance domain knowledge. |
Total Budget (INR): | 24,45,340 |
Organizations involved