Understanding Counting in Small Transformers: The Interplay between Attention and Feed-Forward Layers
Keywords: exploratory analysis, toy transformer, attention, counting, memorization, algorithmic alignment
TL;DR: Different small transformers perform and fail differently on a histogram task depending on their model components alignment with prototypical counting algorithms, their memory capacity and the entanglement of token embeddings.
Abstract: We provide a comprehensive analysis of simple transformer models trained on the histogram task, where the goal is to count the occurrences of each item in the input sequence from a fixed alphabet. Despite its apparent simplicity, this task exhibits a rich phenomenology that allows us to characterize how different architectural components contribute towards the emergence of distinct algorithmic solutions. In particular, we showcase the existence of two qualitatively different mechanisms that implement a solution, relation- and inventory-based counting. Which solution a model can implement depends non-trivially on the precise choice of the attention mechanism, activation function, memorization capacity and the presence of a beginning-of-sequence token. By introspecting learned models on the counting task, we find evidence for the formation of both mechanisms. From a broader perspective, our analysis offers a framework to understand how the interaction of different architectural components of transformer models shapes diverse algorithmic solutions and approximations.
Submission Number: 64
Loading