Explaining Neural Networks (Interview with Joao Leite, NOVA University Lisbon)

Описание: In this interview, we look at a new way to explain the inner workings of a neural network using "mapping neural networks" that uncover concepts used in classification decisions. Joao Leite, a Professor of Computer Science and NOVA University Lisbon describes his work in this area.

0:00:00 Introduction
0:00:51 How Joao Leite started researching the intersection between neural and symbolic AI
0:08:40 Growth of neurosymbolic AI
0:17:54 Symbolic explanation of neural networks
0:25:34 Relevant vs. non relevant concepts
0:28:25 The cost of mapping (why mapping neural networks find concepts faster than training a model on the data)
0:37:05 Pinpointing relevant neurons
0:41:15 Generating explanations
0:45:10 Operating without background knowledge
0:54:00 Future research directions
1:01:00 Advice for researchers

References:

M. de Sousa Ribeiro and J. Leite. Aligning Artificial Neural Networks and Ontologies towards Explainable AI. In Procs. of the Thirty-Fifth AAAI Conference on Artificial Intelligence (AAAI'21), pp. 4932-4940, AAAI Press, 2021.
http://userweb.fct.unl.pt/~jleite/pap...

J. Ferreira, M. de Sousa Ribeiro, R. Gonçalves, J. Leite. Looking Inside the Black-Box: Logic-based Explanations for Neural Networks. In G. Kern-Isberner, G. Lakemeyer, and T. Meyer (eds.), Procs. of the 19th International Conference on Principles of Knowledge Representation and Reasoning (KR'22), 2022.
http://userweb.fct.unl.pt/~jleite/pap...

About the channel:
The Neuro Symbolic Channel provides the tutorials, courses, and research results on one of the most exciting areas in artificial intelligence and machine learning. With content originally from the AI course taught at Arizona State University, this channel brings you the latest at the intersection of symbolic methods (e.g., logic programming) and deep learning. Learn about the latest algorithms, Python packages, and progress toward larger goals such as artificial general intelligence (AGI).