KIND: Knowledge Integration and Diversion for Training Decomposable Models

TL;DR

KIND is a novel pre-training approach that constructs decomposable models by integrating knowledge through SVD, enabling flexible adaptation and transfer in resource-constrained and domain-shifted scenarios.

Contribution

KIND introduces a structural SVD-based pre-training method that creates decomposable models with knowledge components, improving adaptability and transferability.

Findings

Models pre-trained with KIND can be decomposed into learngenes and tailors.

Recombining components enables resource-efficient deployment.

Transferring learngenes mitigates domain shifts effectively.

Abstract

Pre-trained models have become the preferred backbone due to the increasing complexity of model parameters. However, traditional pre-trained models often face deployment challenges due to their fixed sizes, and are prone to negative transfer when discrepancies arise between training tasks and target tasks. To address this, we propose KIND, a novel pre-training method designed to construct decomposable models. KIND integrates knowledge by incorporating Singular Value Decomposition (SVD) as a structural constraint, with each basic component represented as a combination of a column vector, singular value, and row vector from U, \Sigma, and V^\top matrices. These components are categorized into learngenes for encapsulating class-agnostic knowledge and tailors for capturing class-specific knowledge, with knowledge diversion facilitated by a class gate mechanism during training. Extensive experiments demonstrate that models pre-trained with KIND can be decomposed into learngenes and tailors, which can be adaptively recombined for diverse resource-constrained deployments. Moreover, for tasks with large domain shifts, transferring only learngenes with task-agnostic knowledge, when combined with randomly initialized tailors, effectively mitigates domain shifts. Code will be made available at https://github.com/Te4P0t/KIND.