A fragmented neural network ensemble method and its application to image classification
Xu Zhang1, Shuai Liu2, Xueli Wang1
1School of Mathematics and Statistics, Beijing Technology and Business University, Beijing, China.
Scientific Reports
|January 27, 2024
Summary
A new Fragmented Neural Network (FNN) method trains many small neural networks on image pieces. Ensembling these weak networks creates a strong model, achieving high accuracy with reduced complexity and cost.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Deep neural networks are rapidly advancing but are costly and risky to develop.
- Current AI research often prioritizes performance over cost and accessibility.
- Many business applications do not require highly complex AI models.
Purpose of the Study:
- To propose a simple, inexpensive, and accessible AI modeling method for practical applications.
- To introduce the Fragmented Neural Network (FNN) method for efficient AI model development.
- To demonstrate that high accuracy can be achieved without extremely large or complex models.
Main Methods:
- Inspired by random forests, the FNN method randomly samples data and features.
- Images are fragmented into smaller pieces for training weak neural networks.
- Multiple weak neural networks are ensembled using a voting mechanism to form a strong network.
Main Results:
- Experiments on MNIST and CIFAR10 datasets show ensemble accuracy significantly exceeds individual base learners.
- Ensemble network accuracy is comparable to or surpasses full models, with improved robustness.
- The FNN method achieves comparable results to full models using fewer parameters and less data.
Conclusions:
- The Fragmented Neural Network method offers a cost-effective and accessible approach to AI development.
- Ensembling weak learners trained on fragmented data provides a viable alternative to large, complex models.
- This approach enables mass production of AI models through parallel and distributed computing.
