LESSWRONG
LW

Danik

Message

Grokking revisited: reverse engineering grokking modulo addition in LSTM

By Daniil Yurshevich, Nikita Khomich TLDR: we train LSTM model on algorithmic task of modulo addition and observe grokking. We fully reverse engeneer the algorithm learned and propose a way simpler equivalent version of the model that groks as well. Reproducibility statement: all the code is available at the repo....

Dec 16, 2024•4

Danik

Danik — LessWrong

Danik

Message

Grokking revisited: reverse engineering grokking modulo addition in LSTM

Dec 16, 2024•4

Danik

Grokking revisited: reverse engineering grokking modulo addition in LSTM

Nikita Khomich

Nikita Khomich, Danik+ 0 more

Nikita Khomich, Danik

By Daniil Yurshevich, Nikita Khomich

TLDR: we train LSTM model on algorithmic task of modulo addition and observe grokking. We fully reverse engeneer the algorithm learned and propose a way simpler equivalent version of the model that groks as well.

Reproducibility statement: all the code is available at the repo.

Introduction

This post is related to Neel Nanda's post and a detailed description of what grokking is can be found there. The short summary is that grokking is the phenomenon when model when being trained on an algorithmic task of relatively small size initially memorizes the trining set and then suddenly generalizes to the data it hasn't seen before. In our work we train a version... (read 1572 more words →)