Unlike von Neumann processors found in conventional computers, the brain is power efficient, effective at a wide range of computational tasks, and adaptable to new simulations and environments. Artificial Neural Networks (ANNs) are biologically inspired computational systems that mimic the signal processing architecture in the brain and have recently received an explosion of interest. ANNs have dramatically improved speech recognition, visual object recognition, object detection, and genomic reconstruction techniques. Much effort is being put into increasing computational speed and power efficiency of electronic neuromorphic computing; for now, the approaches are however useful only for short timescale applications.
Photonic platforms offer an alternative approach to microelectronics, being potentially able to outperform in computing speed and power efficiency. Indium Phosphide (InP)-based photonics offers a rich platform in terms of amplification, (de)multiplexing, nonlinear effects, and scalability, which are some of the main features requested for obtaining an all Optical Neural Network (ONN). However; crosstalk, noise, and nonlinearities need to be studied for their further development, and critically the tradeoffs between the depth and height for an ONN and the theoretical quality of this ONN must be characterized. Furthermore, the standard application of optical circuits is dominated by “laws of large numbers” due to the massive amounts of photons that are in the signals being used. This implies in particular that noise levels for now are relatively low; however, by going to a regime of low photon density, we expect attractive mathematical properties to emerge in ONNs. In addition, continuous-time learning for optical implementation on chip may be an essential paradigm, but it is currently seen as a major challenge.
|Supervisors||Patty Stabile (TU/e), Jaron Sanders (TU/e)|
Eindhoven University of Technology (TU/e)