Noise-injected Neural Networks for Small Sample Data
Author Information
Author(s): Hua Jianping, Lowey James, Xiong Zixiang, Dougherty Edward R
Primary Institution: Translational Genomics Research Institute
Hypothesis
Can noise-injected neural networks outperform classical neural network designs in small-sample settings?
Conclusion
Noise-injected neural network design is often superior to classical methods and performs competitively with the best alternatives.
Supporting Evidence
- Noise injection can lead to improved performance in neural networks.
- The study found that noise-injected designs often outperform classical neural networks.
- The effectiveness of noise injection varies with the amount of noise and sample size.
Takeaway
This study shows that adding noise to neural networks can help them learn better from small amounts of data, making them more accurate.
Methodology
The study used extensive simulations to compare noise-injected neural networks with classical designs and other classification methods across various small sample sizes.
Limitations
The computational cost of noise injection can be high, especially with many feature sets.
Participant Demographics
The study involved synthetic data and real patient data from breast cancer microarray studies.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website