Inverse Gradient Layer for Multi-Task Learning Forgetting

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing machine learning models struggle to prioritize main tasks over auxiliary tasks during training, leading to suboptimal performance and retention of irrelevant information.

Innovation Solution

A computer-implemented method involving a neural network with multiple heads, where an inverse gradient layer is applied between auxiliary heads and the feature extractor after initial training epochs, allowing the network to forget unhelpful tasks and prioritize main task performance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multi-task learning is performed with multiple heads during training, then the feature extractor benefits from diverse task training, but irrelevant information from auxiliary tasks is retained and harms main task performance

Engineering Contradiction:

Improvemulti-task learning capabilityVSAvoidretention of irrelevant information

Core Design Contradiction:

Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent applies gradient reversal layers that invert the gradient flow during backpropagation. Instead of allowing auxiliary task gradients to directly update feature extractor weights (which causes irrelevant information retention), the reversal layer flips the gradient direction, effectively making the feature extractor learn features that are invariant to auxiliary tasks while still benefiting from multi-task training diversity

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The gradient reversal layer acts as an intermediary component between the auxiliary heads and the feature extractor. It mediates the gradient flow by reversing the gradients before they reach the feature extractor, allowing the system to maintain multi-task learning benefits while preventing harmful gradient propagation that would cause irrelevant information retention

Inventive Principle:

Principle #24Intermediary (Mediator)

2Productivity

If training continues for many epochs with multiple tasks, then the model learns from diverse data, but main task performance deteriorates due to interference from auxiliary tasks

Engineering Contradiction:

Improvetraining efficiencyVSAvoidmain task performance

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

By reversing the gradient direction from auxiliary tasks, the system enables extended training epochs without the usual performance deterioration. The feature extractor continues to learn from diverse auxiliary tasks while the gradient reversal prevents harmful interference, allowing productivity to increase without sacrificing main task reliability

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The gradient reversal mechanism provides a form of feedback control where the system monitors the gradient flow and actively corrects it to prevent performance degradation. This allows the training process to maintain stability over many epochs while continuing to benefit from multi-task learning

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS20250068926A1Forgetfulness mechanism in multi-task learning

Publication Date: 2025.02.27 INTERNATIONAL BUSINESS MACHINE CORPORATION

AI summary

A method, computer system, and a computer program product are provided. A neural network that includes multiple heads is trained. At least one auxiliary head of the multiple heads is identified. After completion of initial epochs of the training, a respective inverse gradient layer between the at least one auxiliary head and a feature extractor of the neural network is applied. Additional epochs of the training with the neural network and the inverse gradient layer are performed.