Open AccessBehavioral Analysis of Neural Network Using Various Visualization Strategies
Author(s) -
Asma Ansari,
Adiba Kalaniya,
Shaziya Memon
Publication year - 2021
Publication title -
international journal of advanced research in science, communication and technology
Language(s) - English
Resource type - Journals
ISSN - 2581-9429
DOI - 10.48175/ijarsct-1118
Subject(s) - visualization , computer science , convolutional neural network , deep learning , artificial intelligence , artificial neural network , machine learning , key (lock) , pattern recognition (psychology) , computer security
Currently extensive researches are focusing on understanding Machine Learning models mainly Deep Learning ones because of their black-box nature. Convolutional neural network (CNN) architectures used in Deep Learning have made their way into computer vision. Saliency maps or attribution maps are mainly used to find the most important features which in turn help us predicting results in the model. In this paper we have worked on different visualization techniques like Gradient Class Activation Map (Grad-CAM), Grad-CAM++, Score-CAM, and Faster Score-CAM on various architectures like VGG16, ResNet-based architectures, and pre-trained models and further investigated their results. Three different datasets (Plant village, Internet augmented, Real-world augmented) have been used and experimented on. The core processes comprise of image capturing, study and implementation of image pre-processing, testing on different neural network architecture, and assessment of data visualization. All of the key steps required to implement the model are detailed throughout the document.