Monitoring-Neuron-Activations

Authors

• Jaswanth Bandlamudi
• Sasi Kiran Gaddipati

Overview

A descriptive visualization of neuron activations for image classification and analyzing the robustness of the classification model by comparing the test image output and corresponding activation patterns. Project link: https://github.com/gsasikiran/Monitoring-Neuron-Activations

Requirements

• Language : Python 3.7

• Coding convention : PEP08

• Testing framework : python unittest

• Data structure:

  • Input : Image and layer number
    - Input the image to get the activation layers

  • Output : Image - Plot the images of the activation layers

• Design Patterns: Bridge pattern

Installation and Running

• Clone the repo

git clone https://github.com/gsasikiran/Monitoring-Neuron-Activations

• Run visualize_activation.py

python visualize_activation.py image_path(with quotes) layer_number

• Example command line

python visualize_activation.py 'images/test_image.png' 2

Description

• The layer number is not zero-based index.

• The 'test_image.png' used for demostration purposes is taken from the triangles dataset[2].
  

Limitations and Future Work

• Currently it works only for 28 x 28 image. This has to be generalized for any image.

• As the number of activation layers, alter for model to model, the visualization of number of activation layers assists in selecting the layer number in index range.

• Monitoring for out of distribution is not achievaible with the implemented approach.

Reference

[1] Cheng, Chih-Hong, Georg Nührenberg, and Hirotoshi Yasuoka. "Runtime monitoring neuron activation patterns." 2019 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 2019.

[2] Azmi, Mohd Sanusi, et al. "Exploiting features from triangle geometry for digit recognition." 2013 International Conference on Control, Decision and Information Technologies (CoDIT). IEEE, 2013.