Multi-Task Active Learning Using Segmentation and Depth Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional active learning techniques are limited in labeling training images capable of improving the performance of multi-task models, particularly in tasks involving image recognition, voice recognition, and natural language processing.
Innovation Solution
An active learning device and method that extracts features from input, segmentation, and depth images using cosine distances to determine and label training images, enhancing the performance of multi-task models by selecting images that minimize cosine distances and balance training across tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional active learning techniques are used, then single-task model performance can be improved, but multi-task model performance cannot be improved
Solution Approach 1:
The patent segments the feature extraction process into three separate feature extractors, each dedicated to a specific task (image recognition, voice recognition, natural language processing). Each feature extractor processes the same input data but produces task-specific features. This segmentation allows the system to handle multiple tasks independently while maintaining specialized processing for each, thereby improving multi-task model performance without compromising single-task performance.
Solution Approach 2:
The patent implements a universal active learning framework that simultaneously serves multiple tasks. The system uses a common set of input data and a unified sampling strategy that considers features from all tasks, making the active learning process multi-functional. This universal approach allows the system to improve performance across multiple tasks using the same active learning mechanism, rather than requiring separate active learning processes for each task.
2Measurement precision
If more training data is labeled to improve model accuracy, then model performance increases, but labeling cost and time increase
Solution Approach 1:
The patent implements a feedback mechanism where the active learning system continuously monitors the performance and uncertainty of the multi-task model. Based on this feedback, the system identifies which unlabeled samples would provide the most valuable information for improving model accuracy. This feedback-driven approach ensures that labeling efforts are focused on samples that will yield the highest return in terms of accuracy improvement, thereby reducing the total number of samples that need to be labeled while still achieving high model performance.
Solution Approach 2:
The patent changes the parameter space by considering multiple task-specific features simultaneously when determining which samples to label. Instead of using a single criterion for sampling, the system evaluates samples based on their importance across multiple tasks and their contribution to reducing overall uncertainty. This multi-dimensional parameter approach allows the system to identify samples that are most valuable for multi-task learning, improving accuracy more efficiently than conventional single-criterion sampling methods.
Data Source
AI summary
An active learning device and an active learning method are provided. The active learning device includes an input device configured to receive an input image, and a controller configured to extract a first feature from the input image, extract a second feature from a segmentation image corresponding to the input image, and extract a third feature from a depth image corresponding to the input image. The active learning device is also configured to determine the input image as a training image based on cosine distances among the first feature, the second feature and the third feature, and labels the training image.