Empirical Capacity Model for Self-Attention Neural Networks

TL;DR

This paper develops an empirical capacity model for self-attention neural networks, enabling better design of models with optimal parameters for specific memorization tasks based on their training algorithms and data.

Contribution

It introduces an empirical capacity model for transformers that links model size and training data to memorization ability, guiding task-specific model design.

Findings

Empirical capacity depends on training algorithms and data content.

The model predicts the memorization capacity of transformers.

Guides optimal model size for specific tasks.

Abstract

Large pretrained self-attention neural networks, or transformers, have been very successful in various tasks recently. The performance of a model on a given task depends on its ability to memorize and generalize the training data. Large transformer models, which may have billions of parameters, in theory have a huge capacity to memorize content. However, the current algorithms for the optimization fall short of the theoretical capacity, and the capacity is also highly dependent on the content. In this paper, we focus on the memory capacity of these models obtained using common training algorithms and synthetic training data. Based on the results, we derive an empirical capacity model (ECM) for a generic transformer. The ECM can be used to design task-specific transformer models with an optimal number of parameters in cases where the target memorization capability of the task can be defined.