Training-free Task Classification for Multi-Task Model Merging

Yeoreum Lee, Jungyong Son, Jinwook Jung, Sungyong Baik

Keywords: model merging, task classification, training-free

Abstract:

The advent of foundation models, coupled with the pretraining-finetuning paradigm, has ignited a proliferation of various tasks and corresponding task-specific models.
This has, in turn, spurred research into finding a unified system that can handle any input coming from various tasks, by combining models via weight interpolation.
While many works have focused on improving the model merging process, most methods assume the knowledge of which task distribution (or task ID) each input belongs to.
While few recent works have attempted to design new merging methods that can handle scenarios where task ID is unknown (task-unknown scenarios), they require either additional training or multiple number of forward passes, undermining the efficiency of a unified framework.
In this work, we aim to empower existing merging methods with the capability of handling task-unknown scenarios, without additional training or multiple number of forward passes.
To this end, we reconceptualize the pursuit of model merging for task-unknown scenarios as a task-classification challenge: identifying the task distribution a given input data belong sto.
Leveraging Gaussian discriminant analysis (GDA), we introduce our method, \textsc{MaD}, which identifies the task identity of input data by comparing the \textbf{Ma}halanobis \textbf{D}istance between input features and each task-conditional Gaussian distribution.
Consequently, \textsc{MaD} can be applied to existing model merging methods in an off-the-shelf manner to empower them with the capability to handle task-unknown scenarios.
Experimental results demonstrate the effectiveness and flexibility of \textsc{MaD} for both computer vision and natural language processing domains, under task-unknown scenarios.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 4359