On-Chip Learning via Transformer In-Context Learning

TL;DR

This paper introduces a neuromorphic transformer model with on-chip plasticity that enables in-context learning and adaptation, reducing memory transfer bottlenecks and highlighting the potential for hardware-efficient, local learning rules.

Contribution

It presents a novel neuromorphic transformer architecture utilizing on-chip plasticity for in-context learning, demonstrating hardware-efficient, local learning rules on Loihi 2.

Findings

Successful demonstration of in-context learning on Loihi 2

Reduced memory transfer through on-chip plasticity

Highlighting the importance of pretrained models for local learning

Abstract

Autoregressive decoder-only transformers have become key components for scalable sequence processing and generation models. However, the transformer's self-attention mechanism requires transferring prior token projections from the main memory at each time step (token), thus severely limiting their performance on conventional processors. Self-attention can be viewed as a dynamic feed-forward layer, whose matrix is input sequence-dependent similarly to the result of local synaptic plasticity. Using this insight, we present a neuromorphic decoder-only transformer model that utilizes an on-chip plasticity processor to compute self-attention. Interestingly, the training of transformers enables them to ``learn'' the input context during inference. We demonstrate this in-context learning ability of transformers on the Loihi 2 processor by solving a few-shot classification problem. With this we emphasize the importance of pretrained models especially their ability to find simple, local, backpropagation free, learning rules enabling on-chip learning and adaptation in a hardware friendly manner.