Locas: Your Models are Principled Initializers of Locally-Supported Parametric Memories

TL;DR

Locas introduces a principled, locally-supported parametric memory for transformers that enables efficient continual learning and memory offloading, improving model generalization and reducing catastrophic forgetting with minimal additional parameters.

Contribution

The paper proposes Locas, a novel parametric memory mechanism for transformers, with a principled initialization method that enhances continual learning and memory integration.

Findings

Locas achieves effective memory storage with only 0.02% additional parameters.

Locas improves generalization and reduces catastrophic forgetting in language modeling.

Locas demonstrates strong performance on long-context dialogue and book modeling tasks.

Abstract

In this paper, we aim to bridge test-time-training with a new type of parametric memory that can be flexibly offloaded from or merged into model parameters. We present Locas, a Locally-Supported parametric memory that shares the design of FFN blocks in modern transformers, allowing it to be flexibly permanentized into the model parameters while supporting efficient continual learning. We discuss two major variants of Locas: one with a conventional two-layer MLP design that has a clearer theoretical guarantee; the other one shares the same GLU-FFN structure with SOTA LLMs, and can be easily attached to existing models for both parameter-efficient and computation-efficient continual learning. Crucially, we show that proper initialization of such low-rank sideway-FFN-style memories -- performed in a principled way by reusing model parameters, activations and/or gradients -- is essential for fast convergence, improved generalization, and catastrophic forgetting prevention. We validate the proposed memory mechanism on the PG-19 whole-book language modeling and LoCoMo long-context dialogue question answering tasks. With only 0.02\% additional parameters in the lowest case, Locas-GLU is capable of storing the information from past context while maintaining a much smaller context window. In addition, we also test the model's general capability loss after memorizing the whole book with Locas, through comparative MMLU evaluation. Results show the promising ability of Locas to permanentize past context into parametric knowledge with minimized catastrophic forgetting of the model's existing internal knowledge.