Active Learning Neural Network for Molecule Energy Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge is to improve the accuracy of a neural network for predicting molecule energy with as few labeled data as possible, given the high computational cost of Density Functional Theory (DFT) calculations.
Innovation Solution
An active learning method that uses a second machine learning model to predict the uncertainty of the first model's predictions, selects unlabeled data with high uncertainty, labels it using DFT calculations, and retrains the first model with the new labeled data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised learning is performed using DFT-calculated labeled data, then prediction accuracy of the neural network is improved, but training cost increases due to the long calculation time of DFT
Solution Approach 1:
The system performs self-service by using its own prediction uncertainty estimates to automatically identify which data points require expensive DFT labeling. The neural network itself generates the selection criteria through uncertainty quantification, eliminating the need for external manual selection processes and reducing overall training time while maintaining accuracy improvement.
Solution Approach 2:
The approach changes the parameter being optimized from pure prediction accuracy to a balance between prediction accuracy and training efficiency. By introducing uncertainty as a selection parameter, the system identifies the most informative data points for labeling, thereby reducing the total number of expensive DFT calculations needed while still achieving accuracy improvements.
2Measurement precision
If more labeled data is collected for supervised learning, then prediction accuracy is improved, but training cost increases due to the computational expense of DFT calculations
Solution Approach 1:
The system autonomously determines which unlabeled data points should be converted to labeled data by evaluating its own prediction uncertainty. This self-service mechanism ensures that DFT calculations are performed only on the most valuable data points, maximizing the impact of each expensive computation and reducing overall computational cost while improving accuracy.
Solution Approach 2:
Instead of labeling all unlabeled data or using a fixed threshold, the system applies partial action by selectively labeling only those data points where uncertainty exceeds a dynamic threshold. This partial labeling approach avoids the excessive computational cost of labeling all data while still achieving sufficient accuracy improvement through the most critical samples.
3Measurement precision
If uncertainty prediction is performed using multiple neural networks, then prediction accuracy is improved, but device complexity increases
Solution Approach 1:
The system merges the primary prediction function and uncertainty estimation function into a unified neural network architecture. By combining these functions, the system achieves accurate uncertainty prediction without requiring separate multiple networks, thereby reducing device complexity while maintaining the ability to identify informative data points for active learning.
Data Source
AI summary
A non-transitory computer-readable recording medium stores an active learning program for causing a computer to execute a process including: predicting, based on a second machine learning model trained by using a set of a plurality of pieces of data and a prediction error of a first machine learning model for each of the plurality of pieces of data as training data, a prediction error of the first machine learning model for each of a plurality of pieces of unlabeled data; generating labeled data by assigning a correct answer label to unlabeled data selected from the plurality of pieces of unlabeled data, based on the predicted prediction error; and retraining the first machine learning model by using the generated labeled data.