deepgeo | first deep neural networks that learn to prove

deepgeo is a Python library. deepgeo has no bugs, it has no vulnerabilities, it has build file available and it has low support. You can download it from GitHub.

We present the first deep neural networks that learn to prove simple Geometry theorems, producing proof steps of both deductive reasoning and creative construction. This contrasts with prior work where reasoning or construction follows hand-designed rules and heuristics. Our models learn from simply observing random sketches of points and lines on a blank canvas, i.e. no example proof designed by human is used. We control the random sketches to verify that the neural nets can indeed solve an unseen problem, as well as problems that require more reasoning steps than the ones they are trained on. Here is the project write-up as a rough description of some concepts used throughout this README and the codebase. Please do not share the report with anyone who does not have access to this Github repository. For the uninitiated in synthetic geometry theorem proof, we suggest having a look at this report on the International Mathematical Olympiad 2018. Both solutions for Problem 1 in this report are perfect examples of elegant and simple geometry proofs. We should pursue to build a system that can attain this level of competence (and hence capable of competing at the IMO). Some nice results in Euclidean geometry includes the Simson Line, or the Butterfly Theorem (Look for proof #14). The most famous theorem includes Morley Trisector, Nine Point Circle, Euler Line, or Feuerbach point . Seeing these theorems can help readers appreciate the beauty of Geometry. This project is at its very early stage where we identify all the key challenges in (1) problem formulation, (2) knowledge representation, (3) modeling, and learning. We then propose a set of solutions to these 3 challenges and built an infrastructure around it. The project is a sanity check if the whole pipeline we imagined as such can work at all.