The Sparse Manifold Transform

on YouTube ↗54m

" summary

Proposes a Sparse Manifold Transform: compute L1 sparse codes of local image patches (e.g. 16×16), learn an embedding of the learned dictionary (hundreds to thousands of atoms) using temporal persistence, then linearly project coefficients to a lower‑dimensional manifold and reconstruct. Implements sparse inference with an LCA/Hopfield‑like circuit (LCA Hopfield) and stacks stages to progressively flatten object manifolds. On video patches the embedding yields smoother, lower‑dimensional trajectories and successful patch reconstructions, so sparse coding + manifold flattening both compresses and preserves recoverable detail.

AI-written summary · about

" chapters (10)

1. 0:00Introduction and motivation
2. 4:09Sparse coding principles
3. 9:04Historical context
4. 13:33Mathematical modeling
5. 21:25Manifolds in visual data
6. 24:59Manifold flattening techniques
7. 29:11Locally linear landmarks
8. 33:48Sparse manifold transform
9. 37:01Results and implications
10. 40:16Q&A and discussion

" related talks

• 25mSimple, Efficient and Neural Algorithms for Sparse CodingAnkur Moitra · Information Theory, Learning and Big Data (Simons Institute, 2015)
• 38mManifold Learning and Dimensionality Reduction for Data VisualizationStefan Kühn · PyData Berlin 2018
• 28mRandomized Constrained Matrix DecompositionsBenjamin Erichson · Randomized Numerical Linear Algebra and Applications 2018
• 58mTensor Decomposition, Sparse Representations and ApplicationsBernard Mourrain · Tensors in Computer Science and Geometry 2014
• 48mInput-sparsity Time Algorithms for Embeddings and Regression ProblemsMichael Mahoney · Workshop on Succinct Data Representations and Applications (Simons Institute Big Data 2013)
• 51mTwo Concise Representations for High-Dimensional Visual DataJohn Wright · Workshop on Succinct Data Representations and Applications (Simons Institute Big Data 2013)
• 26mLearning Sparse Data with Near-optimal Speed and Efficiency from a Variety of Measurement ProcessesSidharth Jaggi · Information Theory, Learning and Big Data (Simons Institute, 2015)
• 29mCompressive SensingIman Mossavat · PyData Eindhoven 2021