Token Turing Machines

TL;DR

Token Turing Machines (TTM) are a novel Transformer-based model with external memory designed for efficient long-sequence visual understanding, outperforming existing models on real-world tasks.

Contribution

The paper introduces Token Turing Machines, a new model combining Transformers with external memory for improved long-sequence processing in visual tasks.

Findings

TTM outperforms other Transformer models on sequential visual tasks.

Efficient processing of long sequences with bounded computational cost.

Model effectively integrates memory for real-world applications.

Abstract

We propose Token Turing Machines (TTM), a sequential, autoregressive Transformer model with memory for real-world sequential visual understanding. Our model is inspired by the seminal Neural Turing Machine, and has an external memory consisting of a set of tokens which summarise the previous history (i.e., frames). This memory is efficiently addressed, read and written using a Transformer as the processing unit/controller at each step. The model's memory module ensures that a new observation will only be processed with the contents of the memory (and not the entire history), meaning that it can efficiently process long sequences with a bounded computational cost at each step. We show that TTM outperforms other alternatives, such as other Transformer models designed for long sequences and recurrent neural networks, on two real-world sequential visual understanding tasks: online temporal activity detection from videos and vision-based robot action policy learning. Code is publicly available at: https://github.com/google-research/scenic/tree/main/scenic/projects/token_turing