Skip to Main Content
 

Global Search Box

 
 
 

ETD Abstract Container

Abstract Header

Hybrid ANN-SNN Co-Training for Object Localization and Image Segmentation

Baltes, Marc Joseph
http://rave.ohiolink.edu/etdc/view?acc_num=ohiou1674142920410841

Abstract Details

2023, Master of Science (MS), Ohio University, Computer Science (Engineering and Technology).
The continued growth of deep learning applications has resulted from the increasing availability of annotated data as well as the advancement of hardware. However, deploying deep learning models on this new hardware results in high energy and computational requirements. A recent development in the field of deep learning has introduced spiking neurons that are used in spiking neural networks (SNNs). These biologically inspired neurons operate on sparse spike trains in order to train and test deep learning models while theoretically consuming less energy when compared to an equivalent ANN model. Different ANN to SNN conversion techniques have been proposed because SNNs can not optimize networks using methods such as backpropagation. In this thesis, we present an intermediate hybrid training step implemented under NengoDL before the ANN is fully converted to an SNN. In this hybrid phase, the forward pass of the network uses spiking activations while the backwards pass switches back to non-spiking activations that are differentiable and are able to be used in backpropagation. Using the spiking activations fine-tunes the connection weights during the hybrid training phase and increases the accuracy of the converted SNN model when compared to a converted SNN without the hybrid training phase. With the new hybrid training scheme, we designed networks and experiments on two applications, object localization and image segmentation. The models were evaluated based on a set of proposed performance metrics. Additionally, the estimated energy consumption for the ANNs and the converted SNNs were compared to provide more information on the energy consumption between ANNs and SNNs. To the best of our knowledge, this is the first implementation of the proposed hybrid training approach that has been tested on these spiking models.
Jundong Liu (Advisor)
Li Xu (Committee Member)
David Chelberg (Committee Member)
David Juedes (Committee Member)
54 p.
Artificial Intelligence; Computer Science
Spiking Neural Network (SNN); Artificial Neural Network (ANN); ANN-to-SNN conversion; detection; segmentation

Recommended Citations

Citations

  • Baltes, M. J. (2023). Hybrid ANN-SNN Co-Training for Object Localization and Image Segmentation [Master's thesis, Ohio University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=ohiou1674142920410841

    APA Style (7th edition)

  • Baltes, Marc. Hybrid ANN-SNN Co-Training for Object Localization and Image Segmentation. 2023. Ohio University, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=ohiou1674142920410841.

    MLA Style (8th edition)

  • Baltes, Marc. "Hybrid ANN-SNN Co-Training for Object Localization and Image Segmentation." Master's thesis, Ohio University, 2023. http://rave.ohiolink.edu/etdc/view?acc_num=ohiou1674142920410841

    Chicago Manual of Style (17th edition)

ohiou1674142920410841
124
© 2023, all rights reserved.
This open access ETD is published by Ohio University and OhioLINK.
Release 3.2.12