mlsquare

mlsquare is an open source, deep tech initiative to democratize ML. The fundamental tenet is, use ML to make ML useful.

Past ML software projects

neural tokenizers to enable the extension of English-centric LLMs to other languages
model grafting via analytical and numerical approximations to enable efficient cross-architecture transfer learning
compute & data efficient training via layer-by-layer and block-by-block LLM training, which is inherently embarrassingly parallel, and with potential analytical update rules (no need for backprop and gradients)
deep kernel machines a framework for composing many transformer-like architectures with specifiable inductive biases
xKANs a framework for composing many KAN-like architectures with specifiable non-parametric function approximators like cubic B-splines, Chebyshev polynomials, Wavelets, etc..
embeddings for tabular data bridge the gap between tabular and modern deep learning models, and make deep learning architectures perform no worse than tree-based models like the battle tested catboost

MLOps: Theory and Practice for upper level undergraduate students
Theory of Deep Learning: Why it works - A Constructionist Approach for upper level undergraduate students
Ilya Sutskever’s 30 papers - Walk through

Yours openly
The Saddle Point

Follow on X/Twitter
Follow on LinkedIn
Join the Discord server here