2. Information processing and computation with photonic reservoir systems
Joni Dambre · 2019
Dambre 2 Information processing and computation with photonic reservoir systems 2.1 Introduction The boundaries of digital computingThe recent surge of research into alternatives for conventional (mostly digital) CMOSbased computing hardware is pushed forward by the limits of scalability.The funda mental assumptions, upon which the success of digital computing is based are mainly its robustness, resulting in extremely low error probabilities, and its inherently low static power consumption, both of which break down for very small devices.The digital computing paradigm and the associated design methodology rely on (almost) error-free operation and cannot deal with inaccurate devices.Already in 1956, Von Neumann discussed how to approximate high-precision dig ital computing with unreliable components [1] by introducing redundancy and ma jority voting.However, the bounds he derives to prove that with enough redundancy, error probabilities can be pushed below any desired threshold, are based on the as sumption that errors occur independently.When the errors are correlated, as is usually the case in real life, an ensemble rarely performs worse than the individual models, but the convergence of the accuracy is no longer guaranteed.In practice, taking an ensemble of unreliable models is now common practice in machine learning.The second most important property to break down is the fact that, using CMOS, we can build implementations of digital gates that consume very little static power.As feature dimensions and isolation layer thickness get thinner, MOSFET transistors and CMOS gates start leaking current in all directions.Whereas the power consumption of a computer was long considered an unimportant issue, it has now become more important than the speed of computation.A few powerful GPUs in the room are an excellent replacement for other heating devices, but our hunger for ever-increasing and ubiquitous computing does not fade, especially in the presence of the huge leaps that are being made with AI and deep learning.Many early attempts of computing with alternative devices stay within the digital model of computation.Clearly, this has its benefits: if you can build gates and flipflops, the whole design methodology can stay in place and the chances of industrial uptake of your new technology improve dramatically.However, thus far, no mapping between inherently analog devices and the basic digital building blocks (binary gates, binary gated memory cells) has been found that can sufficiently outperform transis tors on at least one dimension of performance (size, power consumption per operation or power density, speed) without overly compromising the others and at the same time