The core of the scientific work was performed in WP1-WP4 which is summarised below, while WP5-WP8 covered Recruitment/Management/Governance; Training; Dissemination/Exploitation and Ethics.
WP1: NEW CONCEPTS AND THEORY
There is an urgent need for developing a unified theory of non-digital computing, to integrate the multiplicity of the as yet largely disconnected research in different traditional disciplines. In WP1, a range of architectures, algorithms and analyses were realized by ESRs. These gave new insights into our theoretical understanding of information processing in non-digital physical dynamical systems. The contributions include analyses of information encodings in recurrent neural networks with methods from dynamical systems theory. Deliverables D1.2 and D4.2 give an overview coverage of these lines of work, which was realized in the (often collaborative) PhD projects of the ESRs.
WP2: IMPLEMENTATION AND CHARACTERIZATION
Large-scale photonic neural networks have been designed and tested by several ESRs. This has been possible thanks to the efforts regarding multiplexing strategies and scalability. Photonic neural networks based on optoelectronic systems with delay (time multiplexing), frequency combs (frequency multiplexing), and large-area lasers (space multiplexing) have been demonstrated. Key achievements of WP2 include parallelizing and cascading information processing, and implementing high bandwidth (GHz) systems. Interaction with other WPs has been fruitful. WP2 has successfully implemented concepts designed with WP1 and has significantly improved performance in classification, prediction, and inference tasks. In addition, WP2 has provided design guidelines for photonic integrated circuits demonstrated in WP3. Finally, the promising applications proposed in WP4 have been enabled by the systems developed in WP2.
WP3: INTEGRATED SYSTEMS
In WP3, ESRs have investigated different reservoir architectures (spatial multiplexing, crossbar arrays, frequency multiplexing, etc), both theoretically and experimentally. These have been applied to a number of telecom tasks, mostly dispersion and nonlinearity compensation for different modulation formats (IM, self-coherent, Kramers-Kronig…). Different prototypes have been fabricated and characterized, showing a.o. online learning.
WP4: BENCHMARKING AND APPLICATIONS
WP4 was focused on benchmarking neuromorphic hardware, developing novel applications, and improving hardware performance. Key research objectives included analysing the performance in practical environments, identifying industrial applications, and fostering industry collaborations. The benchmarking aspect of WP4 involved comparing novel neuromorphic solutions with conventional digital solutions. The benchmarking evaluated different computing paradigms, node counts, processing speeds, energy consumption, and footprints. The focus was on analog physical computing and photonic substrates, comparing implementations of reservoir computing, extreme learning machines, and feedforward neural networks.
Detailed results from WP1-4 have been published Open Access (accessed via
https://cordis.europa.eu/project/id/860360/results). A final collaborative review paper on ‘A Perspective on Computing with Physical Substrates’ written by the entire POST-DIGITAL consortium has been accepted for publication in the high-impact journal ‘Reviews in Physics’.
