Task-Incremental Neural Network for Object Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing object detection technologies face challenges in sharing neural network models across multiple tasks, leading to increased network size and computational burden, and are prone to catastrophic forgetting without accessing old task data, particularly in single-stage object detectors.
Innovation Solution
The proposed method introduces a task-incremental learning technique for single-stage object detectors that uses a neural network model with task-specific layers and a polarization mask to determine sub-models, allowing for the reuse of important neural network parameters across tasks, optimizing new tasks without altering the network architecture.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a single-stage object detector shares a neural network model across multiple tasks, then the network size and computational burden are reduced, but catastrophic forgetting occurs when learning new tasks without accessing old task data
Solution Approach 1:
The patent segments the neural network model into task-specific layers and shared layers. Task-specific layers are trained and stored separately for each task, while shared layers are reused across tasks. This segmentation allows the system to maintain a compact shared model structure while preventing catastrophic forgetting through separate task-specific parameter storage.
Solution Approach 2:
The patent changes the parameter representation by storing task-specific parameters separately from shared parameters. Instead of retraining the entire model for each task, the system modifies only the task-specific layer parameters while preserving the shared parameters, thereby avoiding catastrophic forgetting and reducing computational burden.
2Reliability
If task-specific layers are added for each task, then catastrophic forgetting is prevented, but the overall network size and computational burden increase
Solution Approach 1:
The patent merges the task-specific layers with the shared layers in a unified neural network architecture. The task-specific layers are integrated into the existing model structure rather than being completely separate models, allowing parameter sharing and reducing overall network size while still preventing catastrophic forgetting.
Solution Approach 2:
The shared layers in the neural network model serve multiple tasks simultaneously, providing universal functionality across different object detection tasks. This multi-functionality reduces the need for separate full-model copies for each task, thereby reducing network size and computational burden while maintaining the ability to prevent catastrophic forgetting through task-specific parameter management.
3Reliability
If conventional two-stage object detectors are used for task incremental learning, then catastrophic forgetting can be mitigated, but the detection speed is reduced compared to single-stage detectors
Solution Approach 1:
The patent applies dynamic parameter management to the single-stage object detector, allowing the model to adaptively switch between shared and task-specific parameters based on the current task. This dynamic approach maintains the fast detection speed of single-stage detectors while mitigating catastrophic forgetting through intelligent parameter selection and management.
Data Source
AI summary
An object detection method includes the following steps: detecting an environment signal, determining a task mode based on the environment signal, capturing an input image, performing feature extraction on the input image through a sub-model of a neural network model according to the task mode, where the sub-model of the neural network model includes a task-specific layer corresponding to the task mode, where a polarization mask of the task-specific layer determines the sub-model of the neural network model, and outputting an object detection result corresponding to the task mode.