Ingredients for On-Chip Integrated Photonic Neuromorphic Computing
Photonic integrated circuits offer a promising route toward ultrafast, low-latency, and energy-efficient neuromorphic computing by exploiting the intrinsic properties of light propagation. This thesis investigates the fundamental building blocks required to implement neural-network operations directly in integrated photonic hardware, including optical weighting, filtering, routing, summation, and nonlinear transformations. Particular emphasis is placed on recurrent neural architectures based on reservoir computing, a framework especially well suited to photonics because only the output layer needs to be trained.
The thesis first explores frequency-multiplexed photonic reservoirs and demonstrates a programmable silicon photonic spectral filter acting as a trainable readout layer. It then extends these concepts to the time domain through a time-multiplexed photonic readout based on cascaded delay lines and tunable units, enabling independent weighting and optical summation of temporal states. Building on this platform, the thesis introduces an integrated photonic convolutional accelerator for one-dimensional signal processing.
Overall, the results provide a coherent pathway toward scalable photonic neural processors combining wavelength and time multiplexing with programmable optical weighting.
Thesis supervisors: Dr. David Doménech, Dr. Miguel C. Soriano
Jury: Dr. Daniel Pastor Abellán, Dr. Javier Porte Parera, Dra. Silvia Ortín González
This PhD Thesis Defense will be broadcasted in the following zoom link: https://us06web.zoom.us/j/89466064429?pwd=po9p99eAEYVPaNI8xIIGoOIz0hOqaF.1
Detalles de contacto:
Miguel C. Soriano Contact form