Learning Efficient Deep Feature Representations via Transgenerational Genetic Transmission of Environmental Information During Evolutionary Synthesis of Deep Neural Networks

The computational complexity of deep neural networks for extracting deep features is a significant barrier to widespread adoption, particularly for use in embedded devices. One strategy to addressing the complexity issue is the evolutionary deep intelligence framework, which has been demonstrated to...

Full description

Saved in:
Bibliographic Details
Main Authors: Shafiee, M. J., Barshan, E., Li, F., Chwyl, B., Karg, M., Scharfenberger, C., Wong, A.
Format: Conference Proceeding
Language:English
Subjects:
Encoding
Genetics
Neural networks
Probabilistic logic
Stress
Training
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The computational complexity of deep neural networks for extracting deep features is a significant barrier to widespread adoption, particularly for use in embedded devices. One strategy to addressing the complexity issue is the evolutionary deep intelligence framework, which has been demonstrated to enable the synthesis of highly efficient deep neural networks that retain modeling performance. Here, we introduce the notion of trans-generational genetic transmission into the evolutionary deep intelligence framework, where the intra-generational environmental traumatic stresses are imposed to synapses during training to favor the synthesis of more efficient deep neural networks over successive generations. Results demonstrate the efficacy of the proposed framework for synthesizing networks with significant decreases in synapses (e.g., for SVHN dataset, a 230-fold increase in architectural efficiency) while maintaining modeling accuracy and a significantly more efficient feature representation.
ISSN:2473-9944
DOI:10.1109/ICCVW.2017.120