Layer-wise Relevance Propagation (LRP) extended to handle attention layers in Large Language Models (LLMs) and Vision Transformers (ViTs)
Attention-aware LRP (AttnLRP) outperforms gradient- and perturbation-based methods, provides faithful attributions for the entirety of a black-box transformer model while scaling in computational complexitiy $O(1)$ and memory requirements $O(\sqrt{N})$ with respect to the number of layers.
Since we get relevance values for each neuron in the model as a by-product, we know exactly how important each neuron is for the prediction of the model. Combined with Activation Maximization, we can label neurons in LLMs and even steer the generation process of the LLM by activating specialized knowledge neurons in latent space!
For understanding the math behind it, take a look at the ICML 2024 paper!
@InProceedings{pmlr-v235-achtibat24a,
title = {{A}ttn{LRP}: Attention-Aware Layer-Wise Relevance Propagation for Transformers},
author = {Achtibat, Reduan and Hatefi, Sayed Mohammad Vakilzadeh and Dreyer, Maximilian and Jain, Aakriti and Wiegand, Thomas and Lapuschkin, Sebastian and Samek, Wojciech},
booktitle = {Proceedings of the 41st International Conference on Machine Learning},
pages = {135--168},
year = {2024},
editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix},
volume = {235},
series = {Proceedings of Machine Learning Research},
month = {21--27 Jul},
publisher = {PMLR}
}
This project is licensed under the BSD-3 Clause License, which means that LRP is a patented technology that can only be used free of charge for personal and scientific purposes.
⚠️ Because of the high community interest, we release a first version of LXT that might change in the future.
To install directly from PyPI using pip, write:
$ pip install lxt
To reproduce the experiments from the paper, install from a manually cloned repository:
$ git clone https://github.com/rachtibat/LRP-eXplains-Transformers
$ pip install ./LRP-eXplains-Transformers
Tested with transformers==4.44.0
, torch==2.1.0
, python==3.11
Layer-wise Relevance Propagation is a rule-based backpropagation algorithm. This means, that we can implement LRP in a single backward pass! To achieve this, we have implemented custom PyTorch autograd Functions for commonly used operations in transformers. These functions behave identically in the forward pass, but compute LRP attributions in the backward pass. To compute the $\varepsilon$-LRP rule for a linear function $y = W x + b$, you can simply write
import lxt.functional as lf
y = lf.linear_epsilon(x.requires_grad_(), W, b)
y.backward(y)
relevance = x.grad
There are also "super-functions" that wrap an arbitrary nn.Module and compute LRP rules via automatic vector-Jacobian products! These rules are simple to attach to models:
from lxt.core import Composite
import lxt.rules as rules
model = nn.Sequential(
nn.Linear(10, 10),
RootMeanSquareNorm(),
)
Composite({
nn.Linear: rules.EpsilonRule,
RootMeanSquareNorm: rules.IdentityRule,
}).register(model)
print(model)
For a quick demo, we provide modified huggingface source code of popular LLMs, where we already replaced all operations with their equivalent LXT variant.
So that you don't have to edit your source code, we exploited torch.fx to symbolically trace the model and replace layers and functions on-the-fly! This means that many 🤗 models are natively supported. However, not all models are symbolically tracable and gradient-checkpointing doesn't work properly, but we can still use torch.fx as a tool to guide us in the building process!
Applying LRP to new models can be tricky! We provide some basic debugging tools to help you out!
ViTs are vulnerable to noisy attributions, hence we must denoise the heatmap. We propose to use the $\gamma$-LRP rule, where the $\gamma$ parameter must be tuned to the model and dataset!
A documentation of LXT is available, however we will add more in the coming months! The Roadmap lists what is still missing.
Feel free to explore the code and experiment with different datasets and models. We encourage contributions and feedback from the community. We are especially grateful for providing support for new model architectures! 🙏
The code is heavily inspired by Zennit, a tool for LRP attributions in PyTorch using hooks. Zennit is 100% compatible with LXT and offers even more LRP rules 🎉 If you like LXT, consider also liking Zennit ;)