6 ms·
Diagrams are very general and can be extrapolated from quantum field theory to deep learning, see: https://medium.com/@pmigdal/in-the-topic-of-diagrams-i-did-wr
by stared 7y ago
Diagrams are very general and can be extrapolated from quantum field theory to deep learning, see: https://medium.com/@pmigdal/in-the-topic-of-diagrams-i-did-write-a-review-simple-diagrams-of-convoluted-neural-networks-6418a63f9281 https://medium.com/@pmigdal/in-the-topic-of-diagrams-i-did-w...
"In particle physics, people use Feynman diagrams a lot. And these are nothing more or less than a graphical representation of summations and integrations over many variables. Also, as an ex-quantum physicist, I am a big fan of tensor diagrams [...]. You can think about it as the Einstein summation convention with no dummy indices (see Einsum in PyTorch)."
- evanb 7y agoYou're probably better off with some kind of Penrose graphical notation https://en.wikipedia.org/wiki/Penrose_graphical_notation https://en.wikipedia.org/wiki/Penrose_graphical_notation
- galaxyLogic 7y agoor: https://www.quantamagazine.org/physicists-discover-geometry-underlying-particle-physics-20130917/ https://www.quantamagazine.org/physicists-discover-geometry-... “Why are you summing up millions of things when the answer is just one function?”
- stared 7y agoIt’s the same.
- srean 7y agoI was about to mention category theoretic connection but the post mentions Bob so I need not say anymore.
- soVeryTired 7y agoBut... the Einstein summation convention is kind of fine in deep learning. The tensor manipulation in deep learning doesn't involve any of the gnarly contortions that physicists need to do like raising and lowering indices, taking covariant derivatives, or taking traces: and even that stuff moved along just fine without the need for diagramatic representations until the 1960s or so.