graphs - AI?

Molecules — https://github.com/pckroon/pysmiles

1. Mapping molecules to some compute-able form? Macromolecules are large graphs (bonds=edges, nuclei = nodes). I don’t know what kind of operator helps in computing something interesting.
2. Note on that: text data isn’t inherently in numerical form or some entity where you can do some compute
3. converting text to a vector → now you can take dot products, or apply a matrix (which represents some operation) to the vector
4. back to molecules, how do you map them to something in math
5. more suggestions to map everything to a vector is kind-of-boring now, but it could still be powerful, and the innovative idea could be elsewhere in the pipeline

Human pose detection —

1. There are some open source libraries to do this, the results are quite good. It’s based on training datasets like this: http://human-pose.mpi-inf.mpg.de/
2. Here’s a result to see how well this works: https://www.youtube.com/shorts/hl9wo3eiT-Q
3. Sometimes the results are incorrect (like the elbow is off), leg is missing.
4. The human joints are kind of a like a graph. And if there was a way to impose some graph constraints, the result could improve.
5. This was demo’d in CVPR (computer vision pattern recognition conference) 2022 — so things have probably improved. They might have just used a CNN (convolution) for visual identification. Also imposing graph constraints might reduce model size?
6. This kind of work accelerates human survellience :( Also kind of boring, unless it’s really quick to do, which it’s probably not.

Computational graphs in pytorch — https://stackoverflow.com/questions/63582590/why-do-we-call-detach-before-calling-numpy-on-a-pytorch-tensor/63869655#63869655

1. Digging into why pytorch tensors are different than numpy arrays: pytorch tensors have an additional layer that keeps track of the computation that has been done on a particular variable
2. this chain of functions is needed when you use the chain rule to take a derivative on a function that has nested functions
3. they call these chains “computational graph” — it likely becomes complex since these neural nets can be large
4. if any kind of understanding of graphs increases the understanding of neural architectures to reduce model size, or get the model to represent more info, that’s likely useful, and if there is abstraction that might be interesting from a math viewpoint