Speeding Up a Time-Delay Photonic Reservoir

Photonic reservoir computing facilitates high-speed analog information processing, promising significant speed-ups of computationally expensive machine learning tasks. In this work, we use a semiconductor laser with delayed feedback as a time-delay photonic reservoir and train only a linear output layer to solve time series prediction tasks. The reservoir laser is optically driven by an injection laser that encodes the input data using time-multiplexing. Therefore, a randomly drawn step function called mask modulates the input data establishing a neural network of so-called virtual nodes encoded in time [1], [2]. The mask is periodic to the input time
$T$
and has a step duration of
$\theta$
, referred to as node separation. These two timescales and the reservoir delay time
$\tau$
are crucial for the virtual network's computational capabilities [3] and its data rate given by
$1/T$
. Instead of choosing the input time close to the delay time, in this work, we demonstrate in two benchmark tasks that the input time can be set much smaller than the delay time while improving the performance. Additionally, this yields a significant speed-up of the reservoir's data rate even when the delay, e.g., realized by optical fibers, can not be easily reduced.