Invariance, Geometry and Deep Neural Networks
summary
Argues that encoding physics and geometry into networks reduces data needs and increases robustness, using orthonormal latent blocks and manifold/tangent‑space constraints implemented as additional loss terms or modules. Applies these to disentangling autoencoders and LSTM time‑series models on standard benchmarks like NTU and MSR, adding a time‑warping constraint to factor out temporal variability. Reports consistent gains—often 4–8 percentage points on NTU—improved repeatability and stability to input degradation.
AI-written summary · about
chapters (10)
- 0:00Welcome and introduction
- 2:32Guest introduction and background
- 8:48Physics-based principles in computer vision
- 13:05Understanding invariance in vision
- 20:10Manifolds and deep learning constraints
- 26:20Pragmatic deep architectures
- 28:09Experimental framework and results
- 37:20Constraints in loss functions
- 41:32Future of data-constrained learning
- 46:29Closing remarks and thanks
related talks
01h 29mBoot camp on invariances in graph learning
01h 04mRandomized Embeddings and Neural Networks
45mSolving Inverse Problems with Deep Learning
01h 04mNonlinear Model Reduction: Using ML to Enable Rapid Simulation of Extreme-Scale Physics Models
11mGeometric Methods for Machine Learning and Optimization
34mThe physics of deep learning using tensor networks
28mNext-generation frameworks for Large-scale Machine Learning
01h 21mTutorial: Implicit Bias I