The physics of deep learning using tensor networks
summary
Argues that convolutional networks map to tensor networks and that convolutional arithmetic circuits use linear activations with product pooling, not ReLU/max, so nonlinearity comes from products. Connects renormalization-group intuition to learning: restricted Boltzmann machines act like coarse-graining steps and relevant versus irrelevant parameters explain which features survive. Demonstrates bond-dimension as an entanglement bottleneck between image regions and points to a tensor network Python library built on TensorFlow with examples and code.
AI-written summary · about
chapters (18)
- 0:00Physics and machine learning intro
- 1:51Deep learning needs physics
- 3:39Simplicity bias in patterns
- 5:26Physics for scale understanding
- 7:12Renormalization group explained
- 9:19Parameter relevance in physics
- 11:46RBMs and zooming out
- 13:13Physical theories favor simplicity
- 14:18How CNNs work
- 19:16CNNs vs convolutional circuits
- 21:45Tensor basics for ML
- 22:26Tensor networks for CNNs
- 26:24Tensor networks: physics to ML
- 27:48Entanglement and bond dimensions
- 30:18Q&A on fractal broccoli
- 31:22Physics on unsolved DL problems
- 32:47Renormalization group in ML
- 34:13Closing remarks
related talks
01h 01mTensor Networks and Quantum Algorithms
27mDemystifying Deep Neural Networks
28mNext-generation frameworks for Large-scale Machine Learning
41mUnveiling the simple math that powers AI
30mTensor Logic: The Language of AI
34mThe Linear Algebra of Deep Learning
43mThe Linear Algebra of Deep Learning
07mWTT: What the Tensor?