What is Deep Light
DeepLight is a Hellenic Foundation for Research and Innovation (HFRI) and the General Secretariat for Research and Technology (GSRT) three-year project intending to transform photonics into a highly efficient Deep Learning-enabling integrated technology that can operate at >10GHz clock frequencies with sub-pJ/bit energy efficiency values and to deploy Deep Learning algorithms optimally adapted to the idiosyncrasy of the executing photonic neuromorphic hardware through a holistic hardware-software co-design approach.
Motivation
In recent years, Deep Learning (DL) achieved state-of-the-art performance in several challenging tasks, ranging from self-driving cars, and reinforcement learning agents that outperform humans in several games, to medical diagnosis and research. DL models are composed of several non-linear neural network layers, where each of these layers analyzes the output of the previous layer allowing the network to capture increasing complex concepts and model non-linear phenomena as the depth of the network increases. Even though these architectures allowed for tackling several difficult tasks with great success, they come with a significant drawback: tremendous amounts of computing power are needed for training and deploying DL models. Trying to provide a continuous performance improvement in next-generation DL machines has been already identified as a major challenge that can be unlocked only when inspecting closer the inner-anatomy of DL computational setups: a) Energy consumption forms currently the main problem in the computing industry, with interconnects comprising the most consuming sector, b) DL Hardware-software co-design approaches are still in their infancy.
[M. A. Nahmias et al. , "Neuromorphic Photonics," Optics & Photonics News 29(1), 34-41 (2018)]
In this effort, DeepLight aims to design, deploy and experimentally demonstrate a photonic neuromorphic hardware layout exploiting photonic integrated interferometric circuitry and optical memory technology for the weighting and activation functions, respectively. DeepLight intends to design and validate accelerated photonic neuromorphic architectures by introducing WDM and orthogonal IQ modulation formats, yielding a total acceleration factor of more than 2 orders of magnitude compared to state-of-the-art electronic neuromorphic processors. Finally, DeepLight will develop DL algorithms optimally adapted to the peculiarities of the underlying photonic executing infrastructure, establishing the theoretical foundations for optically-enabled Deep Learning models and deploying practical Deep Learning techniques on photonic hardware, with the final aim being to demonstrate well-established datasets executed via photonic circuitry.