Neural Network Training with Domain Expert Supervision
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional pattern recognition methods in medical image analysis require substantial expertise and are cumbersome, involving complex normalization and feature extraction algorithms, while deep learning approaches necessitate large amounts of training data, making them less effective with small datasets.
Innovation Solution
A method that combines traditional pattern recognition with deep learning, using domain expert supervision to identify relevant features and irrelevant variations, allowing a neural network to learn without explicit normalization or feature measurement, thereby mitigating the impact of small training datasets through deep negative and positive supervision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional pattern recognition methods are used, then domain expertise can be incorporated through normalization and feature extraction, but the system becomes complex and computationally costly
Solution Approach 1:
The patent replaces traditional mechanical pattern recognition algorithms (normalization, feature extraction) with a neural network system that learns these transformations automatically through deep learning, substituting manual algorithmic approaches with adaptive neural computation
Solution Approach 2:
The patent introduces an intermediary adversarial network that acts as a mediator to learn and remove irrelevant variations from training data, bridging the gap between raw data and the main prediction model without requiring explicit normalization algorithms
2Extent of automation
If deep learning approaches are used, then automation is improved, but large amounts of training data are required
Solution Approach 1:
The patent extracts and removes irrelevant variations from the training data using an adversarial network, isolating only the relevant features needed for prediction, thereby reducing the effective data requirements while maintaining automation
Solution Approach 2:
The patent performs preliminary processing of training data by learning and removing irrelevant variations before the main training process, preparing the data in advance to reduce the quantity needed for effective model training
3Manufacturing precision
If traditional normalization and feature extraction are performed, then relevant features are emphasized, but the process is cumbersome and requires substantial expertise
Solution Approach 1:
The patent enables the system to perform feature extraction and normalization automatically through neural network learning without requiring manual intervention or expert knowledge, making the process self-service and operationally simple
Solution Approach 2:
The patent transforms static, fixed normalization and feature extraction algorithms into dynamic, adaptive neural network processes that automatically adjust to learn the most relevant features from the data
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
A system comprises processing circuitry configured to perform a training method to train a model for predicting from input data at least one predicted output, the training method comprising: receiving a plurality of training data sets; receiving from a user an identification of a first characteristic of the training data sets as a positive characteristic which is relevant to prediction of the at least one predicted output; receiving from the user an identification of a second characteristic of the training data sets as a negative characteristic which is less relevant or irrelevant to prediction of the at least one predicted output; and training the model, the training of the model comprising: performing supervision of the model using the positive characteristic such that the model is trained to use values for the positive characteristic in the prediction of the at least one predicted output; and performing supervision of the model using the negative characteristic such that the model is trained to discount values for the negative characteristic in the prediction of the at least one predicted output.