Leveraging Neuron Activation Patterns to Explain and Improve Deep Learning Classifiers

Sakib Mostafa, Debajyoti Mondal

Keywords: explainableai, deeplearning, optimization

TL;DR: We propose utilization of the neuron activation pattern using entropy to explain and optimize the performance of a deep learning model.

Abstract: Deep learning models embed all the training information as neuron activation patterns. However, understanding these patterns to improve model performance appears to be a notoriously challenging task. This paper examines the neuron activation patterns of deep learning-based classification models and explores whether model performances can be explained or improved through neurons’ activation behaviour. We first show that the entropy of the neuron activation pattern is related to model performance. We then propose a novel modeling approach that given a trained deep learning model, can leverage the neurons' activation probabilities to further boost the classification accuracy. Our comprehensive experimental study shows notable improvements in classification accuracies (sometimes up to 4.7\%) on benchmark datasets for both classic fully connected neural networks and advanced convolutional neural networks.

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3804