A Network Framework for Small-Sample Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Small-sample learning is crucial for training neural networks with limited data.
• Expanding training datasets improves performance but requires careful constraint management.
• Existing methods may falter due to improper data expansion constraints.

Purpose of the Study:

• To propose conditions for effective incorporation of additional training data.
• To introduce a novel neural network framework for self-training using self-generated data.
• To enhance neural network performance in small-sample learning scenarios.

Main Methods:

• Developed the Small-Sample Learning Network (SSLN) framework.
• SSLN comprises an expression learning network and a sample recall generative network.
• Both networks are built upon the Restricted Boltzmann Machine (RBM) architecture.

Main Results:

• Demonstrated that the SSLN converges comparably to the RBM.
• SSLN showed superior performance over other models on benchmark datasets.
• Experiments were conducted on MNIST Digit, SVHN, CIFAR10, and STL-10 datasets.

Conclusions:

• The proposed SSLN framework effectively addresses challenges in small-sample learning.
• SSLN's self-generated data approach enhances neural network performance.
• The SSLN framework offers a promising solution for data-scarce machine learning tasks.