Analyzing layer norm compositions

[nicholasturner1]

On initial analysis, the layer norm biases seem to have a large portion of the largest terms.
https://github.com/nicholasturner1/gpt-omics/blob/main/notebooks/220509_initial_gptj_run.ipynb
These terms also reach for surprisingly long distances (longer than the attention heads). Could be something interesting here.

[nicholasturner1]

Last week, we realized that the bias terms added to the residual stream by other layers are always passed through a layernorm (with its own bias) before reaching its destination within another layer. Given that, we might instead analyze a fused bias term between each layer instead of splitting them apart.

This is probably a job for another paper.