Time-Spatial Interleaving Photonic Convolutional Accelerators

Tigers Jonuzi

Annual PhD presentation

Convolutional Neural Networks are fundamental machine learning tools for processing image, speech, or audio inputs. The convolutional layer is the core building block of a CNN and is where most of the computation takes place. Here, we propose an integrated photonic convolutional accelerator based on time-space interleaving, using standard generic building blocks to reduce hardware complexity. The architecture is capable of addressing both 2D and 1D convolutional kernels, allowing scalability to more complex networks. In addition, a numerical simulation demonstrates the viability of a supervised online learning algorithm for loading kernel weights in both amplitude and phase while accounting for fabrication tolerances and thermal crosstalk.

Zoom seminar: https://us06web.zoom.us/j/81048596189?pwd=Zo8wuoMXiQM4bBRjMmWHtLwDSwmZHm.1


Contact details:

Miguel C. Soriano

Contact form

This web uses cookies for data collection with a statistical purpose. If you continue browsing, it means acceptance of the installation of the same.

More info I agree