PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.
Here we discuss two different approaches to realize optical neural networks (NN). Part ‘A’ focuses on integrated-photonics solution to realize a perceptron using (i) optoelectronic components to perform the MAC operation and (ii) a more novel concept of an all-optical nonlinearity for thresholding and thermo-optical memory for zero-power synaptic weighting. The latter is unique since the processor’s delay simply depends on the time-of-flight of the photon through the NN, which can be in the picosecond range enabling applications in real-time signal processing. Part ‘B’ discusses our recent process of 4f-based Fourier-optical filtering towards convolution neural networks (CNN); this approach uses massive parallelism (10^6 channels) of digital-mirror-display (DMD) technology. Here the computational expensive convolution is simple realized as a multiplication in the Fourier domain via a lens. Our generation (Gen) 1 system shows 250 TMAC/s process capability at 8-bit resolution.
Volker J. Sorger
"Optical neural networks: from integrated photonics to free-space solutions (Conference Presentation)", Proc. SPIE 11284, Smart Photonic and Optoelectronic Integrated Circuits XXII, 1128406 (9 March 2020); https://doi.org/10.1117/12.2547495
ACCESS THE FULL ARTICLE
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.
The alert did not successfully save. Please try again later.
Volker J. Sorger, "Optical neural networks: from integrated photonics to free-space solutions (Conference Presentation)," Proc. SPIE 11284, Smart Photonic and Optoelectronic Integrated Circuits XXII, 1128406 (9 March 2020); https://doi.org/10.1117/12.2547495