Fusing Model-Based and Function-Based Approaches to AI

Friday, February 22, 2019, 11:00 am - 12:00 pm PDTiCal
This event is open to the public.
AI Seminar
Adnan Darwiche, UCLA Computer Science Department
Video Recording:
https://bluejeans.com/137191985 (Live-stream only, no video-recording)

Neural networks are universal function approximators and have had a great impact on real-world applications when trained from labeled data. Queries on models, such as Bayesian networks, can also be viewed as function approximators when trained discriminately. We show an expressiveness gap between these two types of function approximators and bridge it by introducing “Testing Bayesian Networks.” Queries on these networks synthesize “Testing Arithmetic Circuits (TACs),” which are as expressive as neural networks and can be trained from labeled data using gradient descent. However, TACs can integrate background knowledge and come with some formal guarantees on their behavior that are invariant to how they are trained from labeled data. The motivation of this work is to combine the expressiveness of neural networks with the benefits of models: integrating background knowledge, formal guarantees, less dependence on data, improved robustness and interpretability.

Bio: Adnan Darwiche is a professor and chairman of the computer science department at UCLA. He directs the Automated Reasoning Group, which focuses on symbolic and probabilistic reasoning and their applications, including to machine learning. Professor Darwiche is a AAAI and ACM Fellow and former editor-in-chief of the Journal of Artificial Intelligence Research (JAIR).
« Return to Upcoming Events