Running Programs In Reverse for Deeper A.I.
summary
Claims a compiler, ReverseFlow, that uses parametric inversion plus deep learning to turn a program (e.g., a TensorFlow graph or renderer) into an inverse mapping, even when noninvertible. Demonstrations include inverse kinematics (robot joint angles found from XY targets) and inverse graphics (recovering 3D shape from a 2D image); a Julia-based framework extracts and inverts program primitives and a neural net ranks likely solutions. Results show plausible reconstructions and working demos, though scalability and dead-end parameter regions remain practical limits.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 0:30Mental simulation
- 4:15Mental animation
- 5:49Philosophy
- 13:46Approach
- 17:48Functional Version
- 25:48Julia
- 32:39Conclusion
- 33:15Questions
related talks
10mHow DeepInverse Is Solving Imaging in Science and H...
37mNeural Networks are Programs, Too
39mDifferentiable Functional Programming
34mDifferentiable Functional Programming
36mProbabilistic Programs Which Make (Common) Sense
45mSolving Inverse Problems with Deep Learning
25mInductive Synthesis of Structurally Recursive Functional Programs from Non-recur...
51mDeep Learning - What is it and What It Can Do For You