Neurosymbolic Programming
A key theme in our research is the use of neurosymbolic programs, i.e., models constructed through the composition of neural networks and traditional symbolic code. The neural modules in such a program facilitate efficient learning, while the symbolic components allow the program to use human domain knowledge and also be human-comprehensible. Our research studies a wide variety of challenges in neurosymbolic programming, including the design of language abstractions that allow neural and symbolic modules to interoperate smoothly, methods for analyzing the safety and performance of neurosymbolic programs, and algorithms for learning the structure and parameters of neurosymbolic programs from data.
Selected Publications
Neurosymbolic Reinforcement Learning with Formally Verified Exploration Inproceedings
In: Larochelle, Hugo; Ranzato, Marc'Aurelio; Hadsell, Raia; Balcan, Maria-Florina; Lin, Hsuan-Tien (Ed.): Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.
Learning Differentiable Programs with Admissible Neural Heuristics Inproceedings
In: Larochelle, Hugo; Ranzato, Marc'Aurelio; Hadsell, Raia; Balcan, Maria-Florina; Lin, Hsuan-Tien (Ed.): Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020.
Imitation-Projected Programmatic Reinforcement Learning Inproceedings
In: Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, 8-14 December 2019, Vancouver, BC, Canada, pp. 15726–15737, 2019.
Control Regularization for Reduced Variance Reinforcement Learning Inproceedings
In: Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pp. 1141–1150, 2019.
Neurosymbolic Programming. Foundations and Trends in Programming Languages. Book
2018.
HOUDINI: Lifelong Learning as Program Synthesis Inproceedings
In: Advances in Neural Information Processing Systems 31: Annual Conference on Neural Information Processing Systems 2018, NeurIPS 2018, 3-8 December 2018, Montréal, Canada, pp. 8701–8712, 2018.
