Tangent Model Composition for Ensembling and Continual Fine-tuning

TL;DR

Tangent Model Composition (TMC) enables efficient ensembling and continual learning by combining tangent vectors of models, significantly reducing inference costs and allowing flexible, parallel updates without residual effects.

Contribution

TMC introduces a novel method to compose models via tangent vectors, enabling scalable ensembling and continual fine-tuning without sequential bias or replay buffers.

Findings

TMC improves accuracy by 4.2% over non-linear ensembling.

Reduces inference cost by 2.5x to 10x compared to traditional ensembling.

Outperforms recent continual fine-tuning methods across multiple benchmarks.

Abstract

Tangent Model Composition (TMC) is a method to combine component models independently fine-tuned around a pre-trained point. Component models are tangent vectors to the pre-trained model that can be added, scaled, or subtracted to support incremental learning, ensembling, or unlearning. Component models are composed at inference time via scalar combination, reducing the cost of ensembling to that of a single model. TMC improves accuracy by 4.2% compared to ensembling non-linearly fine-tuned models at a 2.5x to 10x reduction of inference cost, growing linearly with the number of component models. Each component model can be forgotten at zero cost, with no residual effect on the resulting inference. When used for continual fine-tuning, TMC is not constrained by sequential bias and can be executed in parallel on federated data. TMC outperforms recently published continual fine-tuning methods almost uniformly on each setting -- task-incremental, class-incremental, and data-incremental -- on a total of 13 experiments across 3 benchmark datasets, despite not using any replay buffer. TMC is designed for composing models that are local to a pre-trained embedding, but could be extended to more general settings. The code is available at: https://github.com/tianyu139/tangent-model-composition