Hello everyone, I want to share here my small research github repository
I like experimenting with neural network architectures and I’m a self-taught mathematician
My work is related to these two things
pASCNN (p-adic sheaf-coherence neural network), this name is marketing more, the reality is
pASCNN - the experimental research neural network architecture that uses sheaf-theoretic restrictions and a p-adic ultrametric to structure logical inference in complex space
Here the repository: GitHub - kaifczxc-lab/pASCNN: pASCNN: the experimental research neural network architecture what uses sheaf-theoretic restrictions and a p-adic ultrametric to structure logical inference in complex space · GitHub
all code is there (tests, models, results in json)
README includes:
audit, full description about all uniqueness (two readouts, how sheaf/p-adic presented in pASCNN, full PASCNNCell (or just Core) device with visualization and code and basics things, tests comparing pASCNN against a ViT-like transformer
About tests:
Note: Not in all cases, but pASCNN wins by ±2% on average
Various tests were conducted:
MNIST
CIFAR-10 (50/500/1000 examples per class and basic)
Relation OOD benchmarks, including a harder shortcut-reduced setup
Mutation Stability Prediction (bioinformatics, was carried out, it is not in the repository yet)
Tree Benchmark (basic left right structure trees)
Known limitations is the thing it doesn’t have easy way to test it (like extension), but all is not that bad, here have small guide how to test it on another pc in Part 4
Based on the current results, this architecture doesn’t claim to be “better than the MLP/baseline transformer.”
Its unique feature is that, based on various test results, pASCNN is, on average, better than the MLP/baseline transformer for complex structures by +/- 2 percent.
I want to know opinion of knowledgeable people, and how the architecture can be improved
Thanks for reading!