
Representation Learning via Manifold Flattening and Reconstruction
Michael Psenka, Druv Pai, Vishal Raman, Shankar Sastry, Yi Ma
Berkeley Artificial Intelligence Research (BAIR) Lab, May 2023
This work proposes an algorithm for explicitly constructing a pair of neural networks that linearize and reconstruct an embedded submanifold, from finite samples of this manifold. Our suchgenerated neural networks, called flattening networks (FlatNet), are theoretically interpretable, computationally feasible at scale, and generalize well to test data, a balance not typically found in manifoldbased learning methods. We present empirical results and comparisons to other models on synthetic highdimensional manifold data and 2D image data. Our code is publicly available.
SlowDNN 2023, arxiv:2305.01777
