Neuroscience-Informed Interpretability of Intermediate Layers in Artificial Neural Networks

Deep Neural Networks have been successfully implemented in different areas of development and research, including Image Classification, Natural Language Processing, Time-Series Forecasting, and Bioinformatics, among others. However, its complex nature has raised questions about its internal function...

Full description

Saved in:
Bibliographic Details
Main Authors: Pina, Oziel Cortes, Villegas-Jimenez, Adolfo A., Aguilar-Canto, Fernando, Gambino, Omar Juarez, Calvo, Hiram
Format: Conference Proceeding
Language:English
Subjects:
Biology
Convolutional Neural Networks
Decision making
Explainability
Forecasting
Hidden Semantics approach
Natural language processing
Neuroscience
Neuroscience-Informed Interpretability
Semantics
Statistical analysis
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Deep Neural Networks have been successfully implemented in different areas of development and research, including Image Classification, Natural Language Processing, Time-Series Forecasting, and Bioinformatics, among others. However, its complex nature has raised questions about its internal functioning and decision-making, which is critical in different areas. This research seeks to explain the hidden representations of a neural network by using frameworks inspired by Neuroscience, which attempts to understand a very complex neural network, which is the human brain. In this approach, we investigated intermediate and low representation in four different networks: a simple dense Feedforward Neural Network and the Convolutional Neural Networks LeNet-5, VGG-16, and ResNet50, by using similar inputs that were used in Neuroscience experiments. With this framework, we could detect highly selective cells to some of the inputs, remarking some interesting similarities between Biological and Artificial Neural Networks.
ISSN:2161-4407
DOI:10.1109/IJCNN60899.2024.10650110