Training neural networks with end-to-end optical backpropagation
<p>Optical computing is an exciting option for the next generation of machine learning hardware that is fast, parallel and energy efficient. To create a truly all-optical neural network, it is necessary to implement both stages of deployment: inference and training. This in turn requires the a...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Language: | English |
| Published: |
2024
|
| Subjects: |
| _version_ | 1811140009581871104 |
|---|---|
| author | Spall, J |
| author2 | Lvovsky, A |
| author_facet | Lvovsky, A Spall, J |
| author_sort | Spall, J |
| collection | OXFORD |
| description | <p>Optical computing is an exciting option for the next generation of machine learning hardware that is fast, parallel and energy efficient. To create a truly all-optical neural network, it is necessary to implement both stages of deployment: inference and training. This in turn requires the ability to construct multiple linear and nonlinear layers, and implement backpropagation - the primary algorithm for training neural networks - in optics. Training with backpropagation requires information to flow forward and backward through the same network, and imposes conflicting requirements on the mathematical function of the activation layers in each direction. Although a straightforward proposition for a digital processor, implementing these functions in optics has remained elusive, and so prevented any demonstration of true end-to-end optical training to date. This thesis builds on a conceptually-simple scheme to overcome this challenge, to show the first practical demonstration of a multi-layer optical neural network that includes end-to-end optical training. Coherent Fourier optics and spatial light modulation is used to implement the linear layers of a neural network, in the form of optical matrix-vector multiplication with real-valued or complex-valued parameters. The phenomenon of saturable absorption is used to perform the nonlinear neuron activations, and backpropagation is performed optically by means of counter-propagating beams of light, which act analogously to the pump and probe beams of doppler-free saturation spectroscopy. The optical network is used to successfully perform a range of standard benchmark classification tasks, after training the network with a variety of schemes that combine the physical system and a digital model in different ways. In doing so the advantages of hardware-in-the-loop training over traditional <em>in-silico</em> training are shown; improved network accuracy and resilience to errors. This work helps to confirm the potential of building the next generation of hardware for machine learning with analog optics for both inference and training.</p> |
| first_indexed | 2024-09-25T04:15:10Z |
| format | Thesis |
| id | oxford-uuid:3b3db1c2-1278-4aae-91bd-d98066a01b0e |
| institution | University of Oxford |
| language | English |
| last_indexed | 2024-09-25T04:15:10Z |
| publishDate | 2024 |
| record_format | dspace |
| spelling | oxford-uuid:3b3db1c2-1278-4aae-91bd-d98066a01b0e2024-07-18T11:44:26ZTraining neural networks with end-to-end optical backpropagationThesishttp://purl.org/coar/resource_type/c_db06uuid:3b3db1c2-1278-4aae-91bd-d98066a01b0eOpticsNeural networks (computer science)Optical computersEnglishHyrax Deposit2024Spall, JLvovsky, A<p>Optical computing is an exciting option for the next generation of machine learning hardware that is fast, parallel and energy efficient. To create a truly all-optical neural network, it is necessary to implement both stages of deployment: inference and training. This in turn requires the ability to construct multiple linear and nonlinear layers, and implement backpropagation - the primary algorithm for training neural networks - in optics. Training with backpropagation requires information to flow forward and backward through the same network, and imposes conflicting requirements on the mathematical function of the activation layers in each direction. Although a straightforward proposition for a digital processor, implementing these functions in optics has remained elusive, and so prevented any demonstration of true end-to-end optical training to date. This thesis builds on a conceptually-simple scheme to overcome this challenge, to show the first practical demonstration of a multi-layer optical neural network that includes end-to-end optical training. Coherent Fourier optics and spatial light modulation is used to implement the linear layers of a neural network, in the form of optical matrix-vector multiplication with real-valued or complex-valued parameters. The phenomenon of saturable absorption is used to perform the nonlinear neuron activations, and backpropagation is performed optically by means of counter-propagating beams of light, which act analogously to the pump and probe beams of doppler-free saturation spectroscopy. The optical network is used to successfully perform a range of standard benchmark classification tasks, after training the network with a variety of schemes that combine the physical system and a digital model in different ways. In doing so the advantages of hardware-in-the-loop training over traditional <em>in-silico</em> training are shown; improved network accuracy and resilience to errors. This work helps to confirm the potential of building the next generation of hardware for machine learning with analog optics for both inference and training.</p> |
| spellingShingle | Optics Neural networks (computer science) Optical computers Spall, J Training neural networks with end-to-end optical backpropagation |
| title | Training neural networks with end-to-end optical backpropagation |
| title_full | Training neural networks with end-to-end optical backpropagation |
| title_fullStr | Training neural networks with end-to-end optical backpropagation |
| title_full_unstemmed | Training neural networks with end-to-end optical backpropagation |
| title_short | Training neural networks with end-to-end optical backpropagation |
| title_sort | training neural networks with end to end optical backpropagation |
| topic | Optics Neural networks (computer science) Optical computers |
| work_keys_str_mv | AT spallj trainingneuralnetworkswithendtoendopticalbackpropagation |