Please check out the colab notebook for interactive figures and more detailed technical explanations.

This post is part of the work done at Conjecture.

Special thanks to Sid Black, Dan Braun, Carlos Ramón Guevara, Beren Millidge, Chris Scammell, Lee Sharkey, and Lucas Teixeira for feedback on early drafts.

There's a lot of non-linearities floating around in neural networks these days, but one that often gets overlooked is LayerNorm. This is understandable because it's not "supposed" to be doing anything; it was originally introduced to stabilize training. Contemporary attitudes about LayerNorm's computational power range from "it's just normalizing a vector" to "it can do division apparently". And theories of mechanistic interpretability such as features as directions and polytopes are unhelpful, or even harmful, in understanding normalization's impact on a network's representations. After all, normalization doesn't alter the direction of vectors, but it still bends lines and planes (the boundaries of polytopes) out of shape. As it turns out, LayerNorm can be used as a general purpose activation function (you can solve MNIST with a LayerNorm MLP, for example). Concretely, it can do things like this:

and this:

We will explain what's going on in these animations later, but the point is that to develop a strong, principled theory of mechanistic interpretability, we need to grapple with this non-linearity.

In this interactive notebook, we study LayerNorm systematically using math and geometric intuition to characterize the ways in which it can manipulate data. We show that the core non-linearity of LayerNorm can be understood via simple geometric primitives. We explain how these basic primitives may perform semantic operations. For example, folding can be viewed as extracting extremal features (e.g. separating extreme temperatures from normal ones). We leverage these primitives to understand more complex low-dimensional classification tasks with multiple layers of non-linearities. The methods and intuition developed here extend to non-linearities beyond LayerNorm, and we plan to extend them in future work. Below is an interactive summary of the content of the notebook. (If you find the summary interesting, you really should check out the notebook: manipulating the figures helps build geometric intuition.)

Summary

The formula for LayerNorm is something messy like

Lay

...

Posts

Wiki Contributions

Comments