Boundary Optimizing Network (BON) [arXiv Jan 18]

2017, Jan 09    

One Line Summary

• A new approach to generalization for deep neural networks when used for supervised learning.

Motivation

• Despite all the success that deep neural networks have seen in classification tasks, their decision boundaries are not well understood, and are suseptable to adversarial perturbations, so there is need to study their behavior in terms of generalization.

Detailed Summary

• Generator model is trained to augment each data sample until the point of misclassificaiton so that the bounday of the target classifier is more smooth.

Novelty and Contributions

• New approach for making the neural networks more generalizable.
• Prevents overfitting better compared to dropout.

Algorithm

Network

Results

Authors

Marco Singh, Akshay Pai

Sources

Paper