Neural Network Training Using Heterogeneous Data Cohorts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional pattern recognition methods in medical image analysis require substantial expertise for data normalization and feature extraction, while deep learning approaches need large datasets, making them inefficient with small, heterogeneous cohorts of training data.
Innovation Solution
A method that uses deep supervision to train neural networks by identifying relevant features for positive supervision and irrelevant characteristics for negative supervision, reducing the need for explicit normalization and feature extraction, and incorporating domain expert knowledge to enhance performance with limited data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If traditional pattern recognition methods are used with small training datasets, then computational resources and time are reduced, but the system requires substantial domain expertise for data normalization and feature extraction, increasing operational complexity
Solution Approach 1:
The neural network performs self-supervised learning by automatically identifying and extracting relevant features from raw medical images without requiring manual feature engineering or domain expert intervention. The system serves itself by learning representation directly from the data, eliminating the need for external normalization and feature extraction processes.
Solution Approach 2:
The patent replaces the mechanical process of manual feature extraction and data normalization (performed by domain experts) with an automated neural network system. This substitution transforms the workflow from a manual, expertise-dependent process to an automated, data-driven process that learns features directly from raw inputs.
2Extent of automation
If deep learning approaches are used with small, heterogeneous training datasets, then automation is improved, but training efficiency and performance deteriorate due to insufficient data
Solution Approach 1:
The patent applies partial supervision by providing ground truth labels for only a subset of training samples rather than requiring complete labeled datasets. This partial action approach allows the neural network to learn effectively from limited labeled data while using unlabeled data for self-supervised pre-training, improving training efficiency with small datasets.
Solution Approach 2:
The system performs preliminary self-supervised pre-training on available data before fine-tuning with limited labeled examples. This preliminary action of learning general representations from raw data prepares the model for subsequent supervised training, improving overall training efficiency when labeled data is scarce.
3Manufacturing precision
If explicit normalization and feature extraction algorithms are implemented, then manufacturing precision is improved, but device complexity and computational cost increase
Solution Approach 1:
The patent extracts and removes the separate normalization and feature extraction steps from the traditional pipeline, integrating these functions directly into the neural network's learned representation. The network learns to perform normalization and feature extraction implicitly through its training process, eliminating the need for explicit separate algorithms and reducing overall system complexity.
Data Source
AI summary
A system including processing circuitry configured to train a model for predicting from input data at least one predicted output, wherein the processing circuitry is configured to: receive a plurality of training data sets; receive from a user a selection of a first characteristic including positive and negative samples which are relevant variations significant to prediction of the at least one predicted output; receive from the user a selection of a second characteristic including an irrelevant sample which is a spurious variation irrelevant to the prediction of the predicted output; perform positive supervision of the model using the first characteristic such that the training of the model is sensitive to the positive and negative samples of the first characteristic; and perform negative supervision of the model using the second characteristic such that the training of the model is insensitive to the irrelevant sample of the second characteristic.