In-browser Visualizer for Neural Network Training
Author(s)
Dass, Megan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
With the increased use of artificial intelligence (AI) in our everyday lives, there is also a growing interest within the field to truly understand how neural networks come to decisions. Within the computer vision (CV) field specifically, with applications such as facial recognition and object recognition, deep neural networks (DNNs) are commonly used to carry out a variety of CV tasks. However, there remains a need to uncover the blackbox nature of DNNs to improve security, increase public trust in object recognition models, and be able to build more advanced models. We present an open-sourced and in-browser tool that allows users to visualize the inputs of the DNN at various layers and epochs using a dimensionality reduction technique called AlignedUMAP. The 2D dimensionality reduction graph shows users how an input might be classified at various stages in the models training process, as well as allows them to compare different inputs using a spatial visualization to understand how class labels may be closely related.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis