Skip to main content
European Commission logo print header

A NOvel Architecture for a photonics LIquid State machine

Article Category

Article available in the following languages:

All-optical information processing

All-optical information processing

Industrial Technologies icon Industrial Technologies

Reservoir computing represents a new paradigm in information processing, based on the insight that computational power can emerge from a system complexity. The central part of the setup is a vast non-linear network — the reservoir — with nodes needed for information exchange. The connections to the output layer are trained to read the state and map it to the desired output. The EU-funded project 'A novel architecture for a photonics liquid state machine' (NOVALIS) aimed to develop a novel photonic approach to reservoir computing based on an LSM, which is a major type of it. The idea was to replace the network by lasers, acting as nodes. These nodes were highly non-linear to provide the complex dynamics necessary for computations. Implementation of these nodes was achieved by using semiconductor lasers (SLs) with delayed feedback. Optical information injection with 5 Gsamples/s sample rates revealed impressive single SL information-processing capacity. Then, coupling and feedback were established for a two-SL system by using polarisation-maintaining optical fibres. However, scientists could not obtain computation results because of slowly varying modulation at the output intensities. Another implementation of LSMs was a vertical-cavity surface-emitting laser (VCSEL) array that was embedded in a cavity, delay-coupling several laser diodes. Consequently, a complex network was formed, consisting of the connections between individual diodes. This delay network acted as the reservoir. Compared to the previous individually coupled elements, this approach showed major advantages in terms of truly parallel computing, scalability and flexibility. The final project activity for demonstrating parallel information processing was based on spatial light modulators. To avoid the slowly varying amplitude variation, a method based on an all-optical classifier was used to map the liquid state of the VCSEL network. Nevertheless, the detector showed noise levels comparable to the induced transient amplitudes. To demonstrate parallel computing, an all-optical LSM using a single delay-coupled laser reservoir was successfully used. NOVALIS represented a significant effort to implement an all-optical LSM based on multiple laser reservoirs. Overcoming the detector noise problem and implementing the training procedure should establish a novel all-optical, stand-alone machine-learning concept.

Keywords

All-optical, information processing, liquid-state machine, laser, machine-learning, reservoir computing, parallel computing

Discover other articles in the same domain of application