- This event has passed.
Colloquium – Kareem Ahmed
February 28 @ 10:00 am - 11:00 am
Kareem Ahmed
University of California, Los Angeles
February 28, 2024
10:00am
3780 WEB
Neuro-Symbolic AI: A Probabilistic Perspective
Abstract:
Neural networks have been successful across many domains, from functional genomics to path planning. In such domains, we are interested in predicting the value of many outputs that constitute a semantically meaningful object, e.g., a gene sequence. This leads to intractable output spaces for which it is not possible to see enough data to faithfully recover the true function. How then can we hope to develop AI systems that are trustworthy, explainable, and socially aligned? Knowledge is not restricted to data points and is present everywhere! Sets of rules, or constraints, specified in formal language, that characterize the set of predictions admissible in a given problem domain. The challenge, however, is: how do we consolidate the continuous nature of gradient-based learning with the discrete nature of constraints? In this talk, I will construe a neural network as inducing a probability distribution over the space of all possible outputs. I will show that, leveraging domain-expert as well as commonsense symbolic knowledge, fulfilling the desiderata of trustworthiness and explainability reduces to performing probabilistic reasoning on the resulting distribution.
Bio:
Kareem Ahmed is a PhD candidate in Computer Science at the University of California, Los Angeles, where he is advised by Guy Van den Broeck and Kai-Wei Chang. His research aims to exploit the unprecedented capabilities of current neural network architectures to learn from unstructured data while endowing them with logical reasoning capabilities to yield socially aware, trustworthy and explainable systems. His work has been published in top artificial intelligence and machine learning conferences, including AAAI, AISTATS, ICLR, NeurIPS and UAI.
Abstract:
Neural networks have been successful across many domains, from functional genomics to path planning. In such domains, we are interested in predicting the value of many outputs that constitute a semantically meaningful object, e.g., a gene sequence. This leads to intractable output spaces for which it is not possible to see enough data to faithfully recover the true function. How then can we hope to develop AI systems that are trustworthy, explainable, and socially aligned? Knowledge is not restricted to data points and is present everywhere! Sets of rules, or constraints, specified in formal language, that characterize the set of predictions admissible in a given problem domain. The challenge, however, is: how do we consolidate the continuous nature of gradient-based learning with the discrete nature of constraints? In this talk, I will construe a neural network as inducing a probability distribution over the space of all possible outputs. I will show that, leveraging domain-expert as well as commonsense symbolic knowledge, fulfilling the desiderata of trustworthiness and explainability reduces to performing probabilistic reasoning on the resulting distribution.
Bio:
Kareem Ahmed is a PhD candidate in Computer Science at the University of California, Los Angeles, where he is advised by Guy Van den Broeck and Kai-Wei Chang. His research aims to exploit the unprecedented capabilities of current neural network architectures to learn from unstructured data while endowing them with logical reasoning capabilities to yield socially aware, trustworthy and explainable systems. His work has been published in top artificial intelligence and machine learning conferences, including AAAI, AISTATS, ICLR, NeurIPS and UAI.