Multi-Task Learning With Grouped Neck Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Multi-task learning techniques face performance degradation due to negative transfer when learning tasks with different types or opposite gradient directions, leading to increased computational costs and time.
Innovation Solution
A multi-task learning method that groups tasks based on their relationships, configures neck networks for each group, and learns a backbone network along with the neck and head networks using a negative transfer loss function to minimize adverse transfer between tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If multi-task learning is applied to learn multiple tasks simultaneously, then model performance improves when the number of tasks is great and data samples per task are small, but performance degrades when tasks are of opposite or different types due to gradient collision
Solution Approach 1:
The patent segments tasks into multiple groups based on their relationships, with each group having its own neck network. This segmentation prevents gradient collision between tasks of opposite types by isolating them in separate network branches, while still allowing the shared backbone to benefit from multi-task learning. Tasks are grouped according to their gradient compatibility, resolving the contradiction between versatility and reliability.
Solution Approach 2:
The patent introduces neck networks as intermediary components between the shared backbone and task-specific head networks. These neck networks act as mediators that process features for specific task groups, preventing direct gradient conflict between tasks while maintaining the benefits of shared representation learning. The neck networks serve as buffers that isolate incompatible task gradients.
2Reliability
If multiple neck networks are configured for different task groups, then negative transfer between tasks is reduced, but device complexity increases
Solution Approach 1:
The patent divides the network into modular segments: a shared backbone network and multiple independent neck networks, each serving specific task groups. This segmentation reduces negative transfer by isolating task-specific processing paths while maintaining overall system manageability through modular architecture. The head networks remain task-specific, adding minimal complexity.
Solution Approach 2:
The backbone network serves multiple functions by providing shared feature representations to different neck networks, enabling multi-task learning benefits. The neck networks, while specialized for different task groups, share the common backbone infrastructure, reducing overall complexity compared to completely separate models. This multi-functionality approach balances reliability and complexity.
3Productivity
If tasks are grouped and learned with separate neck networks, then computational cost and time are reduced, but the number of network components increases
Solution Approach 1:
The patent merges the backbone network across all task groups, sharing computational resources and feature extractions. This merging reduces overall computational cost and training time compared to separate models, while the added neck networks are relatively lightweight compared to the shared backbone. The combination achieves efficiency without proportional complexity increase.
Solution Approach 2:
The patent segments tasks into groups with dedicated neck networks, allowing parallel processing and more efficient gradient computation. This segmentation enables better computational efficiency by processing related tasks together while isolating conflicting gradients, reducing total training time and computational overhead despite adding network components.
Data Source
AI summary
A multi-task learning method includes determining a relationship between a plurality of tasks, grouping the tasks into at least two groups based on the relationship between the tasks, configuring at least two neck networks respectively corresponding to the at least two groups, and learning a backbone network, the at least two neck networks, and a head network corresponding to each of the tasks based on a loss function for reflecting a negative transfer between head networks included in each group.