Analog Computation with Photonic Spiking Nodes – Anaconda

Artificial neural networks, which are at the heart of recent progress in analog computation and machine learning, are becoming increasingly important for our future data-driven societies. However, these successes rely on the computation power of standard computer architectures, which significantly suffer from poor speed and energy efficiency. In this context, research on neuromorphic hardware attracts more and more attention as it is necessary to develop technological innovative strategies to enhance the computing efficiency of these computing approaches. Our project ANACONDA aims at developing the first large-scale and ultrafast photonic neuromorphic hardwares based on spiking artificial neurons. The systems will be applied to relevant, real-world computational challenges and will potentially allow for important gains in terms of computational speed, parallelism, learning efficiency and energy consumption with respect to current technologies.
Artificial spiking neural networks are composed of nonlinear nodes which take inspiration from biological neurons. Each artificial neuron features multiple inputs and one output, which together define the network’s connection topology. When excited above threshold, neurons respond by a spike in an all-or-nothing fashion, corresponding to a dynamical process called excitability. In this project, we will develop in a first step a large scale opto-electronic test system in view of implementing a full-featured, ultrafast all-photonic spiking neural network. Another major innovation of the project is the fact that we will seek inspiration from the sort of unsupervised learning seen in biological systems that is fundamentally different to the standard supervised deep learning techniques that are currently so dominant. This will allow to mimic the ability seen in humans to learn to respond selectively with just a few repeats of a selected input pattern, using a novel algorithm called JAST.
The project is divided into three scientific workpackages, each addressing a different objective. The first objective is to fabricate and implement large-scale spatiotemporal networks of excitable photonic nodes. We will investigate two complementary systems: one based on spiking opto-electronic oscillators, the other on ultra-fast spiking microlasers. The challenge here will reside in the elaboration of a large array (5x5 and 10x10) of spiking nodes with uniform physical characteristics. The second objective is the implementation of a binary spike signal-coding scheme and the implementation/adjustement of binary-connected networks in the two previous spatio-temporal substrates. Here we will introduce the computational neuroscience concepts. The third objective is to demonstrate the computing capabilities of photonic spiking neural networks. A delay-based architecture with virtual spiking nodes will first be studied using the reservoir computing concept before moving to a full spatiotemporal network. The opto-electronic system will explore advanced network optimization techniques, creating guidance for future realizations of high-performance photonic excitable-system computing. The technologically challenging and more powerful approach of arrays of ultra-fast spiking nodes will be then investigated. Both networks computing abilities will be evaluated based on signal and image recognition tasks, targeting at least state-of-the-art performance.
The consortium gathers three laboratories (C2N, Femto-ST, Cerco) experts in the fields of nanotechnology, photonic neural network computing and computational neurosciences of spiking systems. ANACONDA is thus a multidisciplinary project at the crossroads of different scientific disciplines.
The impact and benefits of ANACONDA will span multiple disciplines of fundamental science, and open the way to far reaching future application in the strategic fields of artificial intelligence and data science.