Jump to Conclusions: Short-Cutting Transformers With Linear Transformations
This work provides a method for elucidating internal model processes, with practical applications in early exit strategies for language models.
The authors tackled the problem of understanding transformer decision-making by approximating hidden representations with linear transformations, achieving up to 7.9% layer savings for GPT-2 and 5.4% for BERT while retaining 95% accuracy.
Transformer-based language models create hidden representations of their inputs at every layer, but only use final-layer representations for prediction. This obscures the internal decision-making process of the model and the utility of its intermediate representations. One way to elucidate this is to cast the hidden representations as final representations, bypassing the transformer computation in-between. In this work, we suggest a simple method for such casting, using linear transformations. This approximation far exceeds the prevailing practice of inspecting hidden representations from all layers, in the space of the final layer. Moreover, in the context of language modeling, our method produces more accurate predictions from hidden layers, across various model scales, architectures, and data distributions. This allows "peeking" into intermediate representations, showing that GPT-2 and BERT often predict the final output already in early layers. We then demonstrate the practicality of our method to recent early exit strategies, showing that when aiming, for example, at retention of 95% accuracy, our approach saves additional 7.9% layers for GPT-2 and 5.4% layers for BERT. Last, we extend our method to linearly approximate sub-modules, finding that attention is most tolerant to this change. Our code and learned mappings are publicly available at https://github.com/sashayd/mat.