Mathematical Reasoning in Latent Space

We design and conduct a simple experiment to study whether neural networks can perform several steps of approximate reasoning in a fixed dimensional latent space. The set of rewrites (i.e. transformations) that can be successfully performed on a statement represents essential semantic features of the statement. We can compress this information by embedding the formula in a vector space, such that the vector associated with a statement can be used to predict whether a statement can be rewritten by other theorems. Predicting the embedding of a formula generated by some rewrite rule is naturally viewed as approximate reasoning in the latent space. In order to measure the effectiveness of this reasoning, we perform approximate deduction sequences in the latent space and use the resulting embedding to inform the semantic features of the corresponding formal statement (which is obtained by performing the corresponding rewrite sequence using real formulas). Our experiments show that graph neural networks can make non-trivial predictions about the rewrite-success of statements, even when they propagate predicted latent representations for several steps. Since our corpus of mathematical formulas includes a wide variety of mathematical disciplines, this experiment is a strong indicator for the feasibility of deduction in latent space in general.

[1]  Sanjit A. Seshia,et al.  Learning Heuristics for Automated Reasoning through Deep Reinforcement Learning , 2018, ArXiv.

[2]  Sarah M. Loos,et al.  HOList: An Environment for Machine Learning of Higher Order Logic Theorem Proving , 2019, ICML.

[3]  Josef Urban,et al.  DeepMath - Deep Sequence Models for Premise Selection , 2016, NIPS.

[4]  Georges Gonthier,et al.  Formal Proof—The Four- Color Theorem , 2008 .

[6]  Cezary Kaliszyk,et al.  Deep Network Guided Proof Search , 2017, LPAR.

[7]  Cezary Kaliszyk,et al.  HOL(y)Hammer: Online ATP Service for HOL Light , 2013, Math. Comput. Sci..

[8]  Honglak Lee,et al.  Action-Conditional Video Prediction using Deep Networks in Atari Games , 2015, NIPS.

[9]  Cezary Kaliszyk,et al.  Can Neural Networks Learn Symbolic Rewriting? , 2019, ArXiv.

[10]  Alexei A. Efros,et al.  Large-Scale Study of Curiosity-Driven Learning , 2018, ICLR.

[11]  Jürgen Schmidhuber,et al.  Recurrent World Models Facilitate Policy Evolution , 2018, NeurIPS.

[12]  Daan Wierstra,et al.  Recurrent Environment Simulators , 2017, ICLR.

[13]  Thibault Gauthier,et al.  Learning to Reason with HOL4 tactics , 2017, ICLP 2017.

[14]  Vladlen Koltun,et al.  Learning to Act by Predicting the Future , 2016, ICLR.

[15]  Sarah M. Loos,et al.  Graph Representations for Higher-Order Logic and Theorem Proving , 2019, AAAI.

[16]  G. G. Stokes.  "J." , 1890, The New Yale Book of Quotations.

[17]  Sarah M. Loos,et al.  Learning to Reason in Large Theories without Imitation , 2019, ArXiv.

[18]  Tobias Nipkow,et al.  A FORMAL PROOF OF THE KEPLER CONJECTURE , 2015, Forum of Mathematics, Pi.

[19]  Sergey Levine,et al.  Model-Based Reinforcement Learning for Atari , 2019, ICLR.

[20]  Jian Wang,et al.  Premise Selection for Theorem Proving by Deep Graph Embedding , 2017, NIPS.

[21]  Roger Wattenhofer,et al.  Using State Predictions for Value Regularization in Curiosity Driven Deep Reinforcement Learning , 2018, 2018 IEEE 30th International Conference on Tools with Artificial Intelligence (ICTAI).

[22]  Thibault Gauthier,et al.  TacticToe: Learning to Reason with HOL4 Tactics , 2017, LPAR.

[23]  Sarah M. Loos,et al.  HOList: An Environment for Machine Learning of Higher-Order Theorem Proving (extended version) , 2019, ArXiv.

[24]  John Harrison,et al.  HOL Light: A Tutorial Introduction , 1996, FMCAD.