Task-Agnostic Vision Transformer for Distributed Learning of Image Processing.

Area of Science:

• Computer Vision
• Machine Learning
• Distributed Systems

Background:

• Traditional distributed learning struggles with clients performing different tasks.
• Vision Transformers excel at learning common representations via global attention in computer vision.

Purpose of the Study:

• To develop a distributed learning framework for multiple, distinct image processing tasks.
• To enable clients to learn unique tasks using only their local data.

Main Methods:

• A novel framework leveraging Vision Transformers for disentangled feature representation.
• Task-specific networks (head/tail) translate local data to a common representation.
• A task-agnostic Vision Transformer on the server learns global attention across client features.
• An alternating training strategy decomposes task-specific and common representations.

Main Results:

• Experimental validation on distributed learning for various image processing tasks.
• Demonstrated synergistic performance improvement for each client using local data.
• Effective disentanglement of local and non-local features.

Conclusions:

• The proposed framework successfully addresses limitations of standard distributed learning for heterogeneous tasks.
• Vision Transformers are effectively adapted for distributed, multi-task learning in image processing.
• The method enhances individual client performance by learning from a shared representation.