Archive / INF Seminars / INF_2022_05_24_Xujie_SI
USI - Email
 
 
Università
della
Svizzera
italiana
INF
 
 
 
   
  main_banner
 

Automated Reasoning with Neuro-Symbolic Learning

 
 
 

Host: Prof. Patrick Eugster

 

Tuesday

24.05

USI Campus EST, room D1.15, Sector D
09:30-10:30
   
 

Xujie Si
McGill University, Canada
Abstract:
Reasoning structured data like programs or unstructured data like images has been a grand challenge. Existing approaches either heavily rely on specialized heuristics designed by human experts or simply exploit large deep neural networks which suffer from many issues like data efficiency, interpretability, lack of formal guarantees, etc. In this talk, I will show how to equip discrete logical reasoning with learning capability through a neuro-symbolic design.

I will first present a differentiable learning and reasoning framework that combines probabilistic reasoning with logical reasoning. This general framework enables learning logical rules for various applications including program analysis and also makes it feasible to reason images jointly with symbolic knowledge base. Then, I will demonstrate a non-differentiable neuro-symbolic framework, based on deep reinforcement learning, for non-trivial reasoning tasks like program verification and synthesis. I will conclude with ongoing work on improving industrial-strength reasoning engines by designing and embedding learnable components, and on improving deep learning through reasoning with proper abstractions.

Biography:
Xujie Si is an Assistant Professor in the School of Computer Science at McGill University. He is also a core academic member at Mila, the Quebec AI Institute, and holds a Canada CIFAR AI Chair. He received his PhD from the University of Pennsylvania in 2020. His research lies in the intersection of programming languages and artificial intelligence. He is broadly interested in developing learning-based techniques to help programmers build better software with less effort, integrating logic programming with differentiable learning systems for interpretable and scalable reasoning, and leveraging programming abstractions for data-efficient learning. His work has been recognized with ACM SIGPLAN distinguished paper award and several spotlights in top programming languages and machine learning venues.