This work was produced at Apollo Research, based on initial research done at MATS.

LayerNorm is annoying for mechanstic interpretability research (“[...] reason #78 for why interpretability researchers hate LayerNorm” – Anthropic, 2023).

Here's a Hugging Face link to a GPT2-small model without any LayerNorm.

The final model is only slightly worse than a GPT2 with LayerNorm[1]:

DatasetOriginal GPT2Fine-tuned GPT2 with LayerNormFine-tuned GPT without LayerNormOpenWebText (ce_loss)3.0952.9893.014 (+0.025)ThePile (ce_loss)2.8562.8802.926 (+0.046)HellaSwag (accuracy)29.56%29.82%29.54%I fine-tuned GPT2-small on OpenWebText while slowly removing its LayerNorm layers, waiting for the loss to go back down after reach removal:

Introduction

LayerNorm (LN) is a component in Transformer models that normalizes embedding vectors to have constant length; specifically it divides the embeddings by their standard deviation taken over the hidden dimension. It was originally introduced to stabilize and speed up training of models (as a replacement for batch normalization). It is active during training and inference.

_mathrm{LN}(x) = frac{x - [...] ---

Outline:

(01:11) Introduction

(02:45) Motivation

(03:33) Method

(09:15) Implementation

(10:40) Results

(13:59) Residual stream norms

(14:32) Discussion

(14:35) Faithfulness to the original model

(15:45) Does the noLN model generalize worse?

(16:13) Appendix

(16:16) Representing the no-LayerNorm model in GPT2LMHeadModel

(18:08) Which order to remove LayerNorms in

(19:28) Which kinds of LayerNorms to remove first

(20:29) Which layer to remove LayerNorms in first

(21:13) Data-reuse and seeds

(21:35) Infohazards

(21:58) Acknowledgements

The original text contained 4 footnotes which were omitted from this narration.

The original text contained 5 images which were described by AI.

---

First published:
August 8th, 2024

Source:
https://www.lesswrong.com/posts/THzcKKQd4oWkg4dSP/you-can-remove-gpt2-s-layernorm-by-fine-tuning-for-an-hour

---

Narrated by TYPE III AUDIO.

---

Images from the article: