GenSLMs are a LLM, but genome sequence

1. Take genome sequence
2. Throw transformers at it
3. create “semantic embedding”
4. autoregression happens

This is trained as a foundational model to organize the genomic sequence

Turns out, the embedding space above can be used to discover relations. See proteins can be encoded as hierarchies

A Geometric Brownian Motion is a Brownian Motion with a drift.

It is determined by:

\begin{equation} \dd{S_{t}} = \mu S_{t} \dd{t} + \sigma \dd{S_{t}} \dd{W_{t}} \end{equation}

where, \(S_{t}\) is a Geometric Brownian Motion, \(\mu\) is its drift, \(\sigma\) the volatility, and \(W_{t}\) a centered Brownian Motion.

The geometric multplicity for a given eigenvalue is the dimension of its generated eigenspace.

(Py)Torch is a great C++/Python library to construct and train complex neural networks. It has taken over academia over the last few years and is slowly taking over industry. Let’s learn about how it works!

This document is meant to be read cover-to-cover. It makes NO SENSE unless read like that. I focus on building intuition about why PyTorch works, so we will be writing unorthodox code until the very end where we put all ideas together.