Discrimination Model Training Data Sufficiency Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for preventing excessive adaptation in data recognition tasks, such as overlearning, face challenges when dealing with insufficient training data, leading to poor generalization and classification errors, and existing countermeasures like reducing model parameters or early stopping are not always effective or user-friendly.
Innovation Solution
An information processing apparatus and method that detects the possibility of excessive adaptation by analyzing the number of training images and notifies the user to add more data, incorporating a decision unit to assess the sufficiency of training data and a notification unit to prompt data addition, thereby preventing overadaptation and improving classification precision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the number of training data is small, then the learning speed is fast and the model complexity is low, but the model cannot be adapted to unknown data and general-purpose performance is insufficient
Solution Approach 1:
The patent applies preliminary action by generating synthetic training data before the actual learning process. The distribution generation unit creates artificial training data based on estimated distributions from initial training data, and the distribution estimation unit prepares these distributions in advance. This preliminary generation of synthetic data ensures that when the model learns, it has sufficient diverse training examples, improving general-purpose performance without requiring large amounts of real training data.
2Reliability
If the model parameters are reduced to avoid excessive adaptation, then the model complexity is lowered, but the model cannot correctly learn from the training data
Solution Approach 1:
The patent applies copying by creating synthetic copies of training data through the distribution generation unit. Instead of reducing model parameters, the system generates multiple synthetic training examples that copy the underlying data distribution. This allows the model to maintain its full parameter set and learning accuracy while training on augmented synthetic data, avoiding excessive adaptation to limited real data.
3Reliability
If early stopping of modeling calculation is implemented, then the risk of excessive adaptation is reduced, but the adaptation to training data becomes insufficient
Solution Approach 1:
The patent applies preliminary action by pre-generating synthetic training data and estimating data distributions before the main learning process. The distribution estimation unit and distribution generation unit prepare synthetic training examples in advance, allowing the model to continue learning for the full training period without early stopping. This preliminary preparation ensures the model can learn from diverse synthetic data, improving generalization ability while maintaining sufficient learning time.
Data Source
AI summary
The presence of possibility of occurrence of an excessive adaptation due to use of only learned training data is detected during a learning stage. The user is urged to add data and other information, thereby avoiding the excessive adaptation. For this purpose, the invention has: an inputting unit for inputting a learning image; a generating unit for generating a discrimination model used to decide whether or not a target is normal on the basis of the learning image; a deciding unit for deciding whether or not the number of input learning images is insufficient when the discrimination model is generated; and a notifying unit for notifying a message for urging the user to add the learning image when it is decided that the number of input learning images is insufficient.