SATNet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver

Integrating logical reasoning within deep learning architectures has been a major goal of modern AI systems. In this paper, we propose a new direction toward this goal by introducing a differentiable (smoothed) maximum satisfiability (MAXSAT) solver that can be integrated into the loop of larger deep learning systems. Our (approximate) solver is based upon a fast coordinate descent approach to solving the semidefinite program (SDP) associated with the MAXSAT problem. We show how to analytically differentiate through the solution to this SDP and efficiently solve the associated backward pass. We demonstrate that by integrating this solver into end-to-end learning systems, we can learn the logical structure of challenging problems in a minimally supervised fashion. In particular, we show that we can learn the parity function using single-bit supervision (a traditionally hard task for deep networks) and learn how to play 9x9 Sudoku solely from examples. We also solve a "visual Sudok" problem that maps images of Sudoku puzzles to their associated logical solutions by combining our MAXSAT solver with a traditional convolutional architecture. Our approach thus shows promise in integrating logical structures within deep learning.

[1]  J. Zico Kolter,et al.  OptNet: Differentiable Optimization as a Layer in Neural Networks , 2017, ICML.

[2]  Milind Tambe,et al.  Melding the Data-Decisions Pipeline: Decision-Focused Learning for Combinatorial Optimization , 2018, AAAI.

[3]  Yoshua Bengio,et al.  Gradient-based learning applied to document recognition , 1998, Proc. IEEE.

[4]  J. Z. Kolter,et al.  The Mixing method: low-rank coordinate descent for semidefinite programming with diagonal constraints , 2017 .

[5]  Willem Jan van Hoeve,et al.  The Power of Semidefinite Programming Relaxations for MAX-SAT , 2006, CPAIOR.

[6]  Fan Yang,et al.  Differentiable Learning of Logical Rules for Knowledge Base Reasoning , 2017, NIPS.

[7]  Andreas Krause,et al.  Differentiable Learning of Submodular Models , 2017, NIPS 2017.

[8]  Guy Van den Broeck,et al.  A Semantic Loss Function for Deep Learning with Symbolic Knowledge , 2017, ICML.

[9]  Ole Winther,et al.  Recurrent Relational Networks , 2017, NeurIPS.

[10]  Richard Evans,et al.  Learning Explanatory Rules from Noisy Data , 2017, J. Artif. Intell. Res..

[11]  V. 2Rk⇥n.  SATNet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver , 2019 .

[12]  Ole Winther,et al.  Recurrent Relational Networks for Complex Relational Reasoning , 2018, ArXiv.

[13]  Priya L. Donti,et al.  Task-based End-to-end Model Learning in Stochastic Optimization , 2017, NIPS.

[14]  Alexander I. Barvinok,et al.  Problems of distance geometry and convex properties of quadratic maps , 1995, Discret. Comput. Geom..

[15]  J. Zico Kolter,et al.  What game are we playing? End-to-end learning in normal and extensive form games , 2018, IJCAI.

[16]  J. Zico Kolter,et al.  Low-rank semidefinite programming for the MAX2SAT problem , 2019, AAAI.

[17]  Ohad Shamir,et al.  Failures of Gradient-Based Deep Learning , 2017, ICML.

[18]  Jimmy Ba,et al.  Adam: A Method for Stochastic Optimization , 2014, ICLR.

[19]  Eric P. Xing,et al.  Harnessing Deep Neural Networks with Logic Rules , 2016, ACL.

[20]  Alessandra Russo,et al.  DeepLogic: End-to-End Logical Reasoning , 2018, ArXiv.

[21]  Luc De Raedt,et al.  Neural-Symbolic Learning and Reasoning: Contributions and Challenges , 2015, AAAI Spring Symposia.

[22]  David P. Williamson,et al.  Improved approximation algorithms for maximum cut and satisfiability problems using semidefinite programming , 1995, JACM.

[23]  Alessandra Russo,et al.  DeepLogic: Towards End-to-End Differentiable Logical Reasoning , 2018, AAAI Spring Symposium: Combining Machine Learning with Knowledge Engineering.

[24]  J. Zico Kolter,et al.  The Mixing method: coordinate descent for low-rank semidefinite programming , 2017, ArXiv.

[25]  Andreas Krause,et al.  Differentiable Submodular Maximization , 2018, IJCAI.

[26]  Steven Schockaert,et al.  Lifted Relational Neural Networks: Efficient Learning of Latent Relational Structures , 2018, J. Artif. Intell. Res..

[27]  Simon Haykin,et al.  GradientBased Learning Applied to Document Recognition , 2001 .

[28]  Luc De Raedt,et al.  DeepProbLog: Neural Probabilistic Logic Programming , 2018, BNAIC/BENELEARN.

[29]  Yang Yu,et al.  Tunneling Neural Perception and Logic Reasoning through Abductive Learning , 2018, ArXiv.

[30]  Gábor Pataki,et al.  On the Rank of Extreme Matrices in Semidefinite Programs and the Multiplicity of Optimal Eigenvalues , 1998, Math. Oper. Res..

[31]  David L. Dill,et al.  Learning a SAT Solver from Single-Bit Supervision , 2018, ICLR.