Neural Network Feature Addition for Multi-Task Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning technologies face challenges in efficiently adding new classification tasks to a learned neural network without compromising performance on existing tasks, particularly in medical image segmentation where data scarcity hampers the ability to maintain recognition performance across multiple organs.
Innovation Solution
A machine learning method that involves adding new feature amounts to intermediate layers of a pre-trained neural network, generating a new neural network structure that maintains the existing network's architecture, and optimizing the new feature amounts through learning, allowing for the addition of new classification tasks without altering the output performance of the existing tasks.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If independent neural networks are used for each organ classification task, then recognition performance for specific organs can be optimized, but calculation complexity increases and execution speed decreases when multiple organs need to be classified
Solution Approach 1:
The patent combines multiple independent neural networks into a single integrated network that shares common feature extraction layers. The network uses a shared encoder for feature extraction and separate decoders for different organ classifications, allowing multiple organs to be classified simultaneously with reduced calculation complexity while maintaining recognition performance.
Solution Approach 2:
The patent creates a universal neural network architecture that can perform multiple organ classification tasks through a single model. The shared feature extraction components serve multiple classification objectives, enabling the system to classify different organs using the same network without requiring separate independent networks for each organ.
2Measurement precision
If a neural network is re-trained with new learning data to add classification for new organs, then recognition performance for new organs improves, but output performance for existing organs deteriorates
Solution Approach 1:
The patent performs preliminary training on existing organ classification tasks before adding new classification capabilities. The shared feature extraction layers are pre-trained on existing data, and when new organs need to be classified, only the new decoder branches are trained while the shared encoder parameters remain fixed or are fine-tuned minimally, preventing degradation of existing classification performance.
Solution Approach 2:
The patent segments the neural network into independent functional modules: a shared encoder for feature extraction and separate decoder branches for different organ classifications. This segmentation allows new classification tasks to be added by introducing new decoder branches without modifying the pre-trained encoder, thus maintaining stability of existing classification performance while enabling new capabilities.
3Measurement precision
If multiple independent neural networks are trained separately for different organs, then each network can be optimized for its specific task, but the overall system requires more calculation resources and time
Solution Approach 1:
The patent merges multiple task-specific neural networks into a single multi-task network that shares common computational resources. The shared encoder processes input images once and generates features that are then used by multiple decoder branches for different organ classifications, eliminating redundant calculations and improving execution speed while maintaining task-specific optimization through separate decoder modules.
Data Source
AI summary
A machine learning method and an apparatus, a program, a learned model, and a discrimination apparatus capable of controlling a calculation amount by learning a new task without changing output performance for an existing task in a learned network are provided. A machine learning method according to one aspect of the present disclosure includes a step of adding a new feature amount to at least one intermediate layer included in a learned first neural network that has learned a task of performing first class classification, a step of generating a second neural network having a structure in which a network structure of a calculation path of an existing feature amount of the first neural network is maintained and a new feature amount of a subsequent layer is calculated by performing processing of convolving each of the existing feature amount and the new feature amount, and a step of causing the second neural network to acquire a processing function of performing second class classification by performing learning of the second neural network using a set of learning data corresponding to the second class classification.