Neuro-symbolic artificial intelligence is an emerging field of research that brings together two historically distinct paradigms: machine learning and artificial neural networks (connectionist approach) and knowledge representation and reasoning(symbolic approach). The so-called “third wave” of AI aims at integrating these two paradigms in order to achieve more robust, flexible and human-interpretable systems. The course will first provide an introduction to the neuro-symbolic methods and paradigms that have emerged in the past years, and different ways that knowledge representation and reasoning can be integrated with data-driven machine learning systems. Secondly, it will provide theoretical and practical knowledge of logic tensor networks, a neuro-symbolic framework that uses a differentiable first-order logic language to integrate data-driven examples and logical axioms. It will be shown how logic tensor networks can be used to solve tasks such as classification, clustering and image interpretation tasks using the satisfaction of symbolic rules as an objective. Finally, an overview of neuro-symbolic methods for vision and language applications will be presented.

Knowledge about machine and deep learning; some experience with Python programming and deep learning frameworks

- Strengths and weakness of connectionist and symbolic approaches - Introduction to neuro-symbolic methods and frameworks, composition patterns - Introduction to real logic and logic tensor networks - Symbol grounding in logic tensor networks - Learning as a best satisfiability problem - Practical examples (MNIST classification, clustering, object detection) - Neuro-symbolic methods in semantic image interpretation: scene graph generation, visual query answering - Integration of prior knowledge in deep learning models - Neuro-symbolic methods for explainable AI

In presenza

Presentazione orale

P.D.1-1 - Gennaio