Adaptive Feature Distillation for Negative Transfer in Multi-Task Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multi-task learning techniques in computer vision and image processing often suffer from performance degradation in more difficult tasks due to imbalances in task difficulty, leading to negative transfer. Existing solutions fail to preserve the storage and efficiency of multi-task neural network frameworks.
Innovation Solution
The method employs a multi-task neural network framework trained using an adaptive feature distillation function. This function compares intermediate features of the multi-task network with those of single-task networks, aligning the feature spaces to improve performance while maintaining efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multi-task learning uses a shared framework with large proportion of shared parameters, then storage efficiency is improved and inference speed is improved, but performance degradation occurs on some tasks due to negative transfer
Solution Approach 1:
The patent segments the feature extraction process by introducing task-specific branches after the shared backbone. Each branch independently processes features for its specific task, preventing negative transfer while maintaining the efficiency of the shared framework. The segmentation allows each task to have dedicated parameter sets that do not interfere with other tasks.
Solution Approach 2:
The patent introduces an intermediary mechanism (task-specific branches with selective parameter sharing) between the shared backbone and task outputs. This intermediary layer allows the system to benefit from shared feature extraction while preventing performance degradation through task-specific adaptations.
2Quantity of substance
If multi-task learning uses a shared framework, then storage requirements are reduced, but performance gains in some tasks come at the cost of performance degradation on other tasks
Solution Approach 1:
The patent applies local quality by making parameters local (task-specific) where needed while keeping them shared where beneficial. Task-specific branches have dedicated parameters for tasks that are sensitive to negative transfer, while the backbone shares parameters for common feature extraction, optimizing both storage and performance.
Solution Approach 2:
The patent introduces dynamic parameter sharing where the degree of parameter sharing is adjusted based on task characteristics. Some parameters are shared across tasks while others are task-specific, creating a dynamic architecture that adapts to different task requirements and prevents negative transfer.
Data Source
AI summary
A method, system, apparatus, and non-transitory computer-readable medium for image processing using a multi-task neural network framework may be provided. The method be executed by one or more processors, and may include receiving an input image; and generating one or more image patches based on the input image. The method may include performing an image processing task based on the input image using the multi-task neural network framework, wherein the multi-task neural network framework is trained using an adaptive feature distillation function, and wherein the adaptive feature distillation function is based on a comparison of intermediate features of the multi-task neural network framework and intermediate features of a plurality of single-task neural network models. The method may include generating an output of the image processing task based on up sampling an output of the multi-task neural network framework.