Multi-Task Learning Device Using Feature Loss for Heterogeneous Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-task learning schemes require large-scale labeled datasets for each task, leading to high costs and time consumption, especially when adding new tasks, and are inefficient under limited hardware resources.
Innovation Solution
A multi-task learning device and method that utilizes a feature extraction layer, decoding layers, and a parameter updater to generate and update features and inference results across heterogeneous datasets, incorporating a feature loss layer to adapt parameters for multiple tasks, reducing the need for extensive labeling and resource-intensive data construction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multi-task learning is performed using a single public network and multiple output layers, then multiple tasks can be processed in real time under limited hardware resources, but a large-scale data set including labels corresponding to all tasks is required for each training data, leading to huge time and money investment
Solution Approach 1:
The patent applies universality by using a single public network to perform multiple tasks simultaneously through multiple output layers. The network is designed to handle diverse tasks (e.g., segmentation, detection, depth estimation) without requiring separate networks for each task, thereby achieving multi-functionality under limited hardware resources while reducing overall system complexity
Solution Approach 2:
The patent applies preliminary action by pre-processing training data to generate pseudo-labels automatically without requiring manual labeling for all tasks. The system performs preliminary feature extraction and uses confidence thresholds to automatically label data, significantly reducing the time and cost of data set construction before the actual multi-task learning process begins
2Reliability
If labeling is performed for existing tasks, then supervised learning can be conducted, but when learning on a new task is required, a labeling process corresponding to the new task is additionally required for all data sets, increasing cost and time
Solution Approach 1:
The patent applies self-service by enabling the system to automatically generate pseudo-labels for new tasks without requiring external manual labeling. When a new task is introduced, the system uses the existing public network to process training data, generates predictions, and automatically creates labels based on confidence thresholds. This self-labeling mechanism allows rapid adaptation to new tasks while maintaining learning accuracy
Solution Approach 2:
The patent applies preliminary action by pre-establishing a framework that can automatically adapt to new tasks. The system pre-configures multiple output layers and loss functions that can be activated for new tasks without requiring complete re-labeling of datasets. The preliminary setup includes automatic pseudo-label generation mechanisms that prepare the system for rapid task adaptation
3Adaptability or versatility
If heterogeneous data sets with different label types are used for multi-task learning, then task-specific learning can be achieved, but the complexity of constructing and managing diverse labeled data increases
Solution Approach 1:
The patent applies universality by designing a unified public network that can process heterogeneous data sets with different label types through multiple output layers. Each output layer is configured for a specific task (e.g., segmentation, detection, depth estimation), but they all share the same input processing and feature extraction pathways, thereby supporting task diversity while maintaining data construction simplicity
Solution Approach 2:
The patent applies preliminary action by pre-processing heterogeneous training data to a unified format before input to the public network. The system performs preliminary operations such as image normalization, feature extraction, and automatic pseudo-label generation that standardize diverse data sets, reducing the complexity of constructing and managing heterogeneous labeled data while maintaining task-specific learning capabilities
Data Source
AI summary
A multi-task learning device includes a feature extraction layer that generates a first feature corresponding to a first image and a second feature corresponding to a second image; a first decoding layer that generates a first task inference result corresponding to the first image; a second decoding layer that generates a second task inference result corresponding to the second image; a first loss layer that generates a first task loss with reference to the first task inference result and a first task ground truth (GT) result corresponding to the first task inference result; a second loss layer that generates a second task loss with reference to the second task inference result and a second task GT result corresponding to the second task inference result; a feature loss layer that generates a feature loss with reference to the first feature and the second feature; and a parameter updater that updates parameters of at least some of the various layers.