Martin Trefzer - Reservoir Computing with Physical Substrates скачать в хорошем качестве

Martin Trefzer - Reservoir Computing with Physical Substrates

Date: 5/11/21 Martin Trefzer, University of York Title: Reservoir Computing with Physical Substrates Abstract: Classical digital computing is power hungry, fragile, and hard to interface to the real world. Unconventional computers can help overcome these issues, particularly by being able to perform computation embodied in physical materials that can directly exploit the natural dynamics of the substrate, thereby dramatically reducing the power requirements. Reservoir computing (RC) is a computational model derived from recurrent neural network theory and has proven to be an efficient approach for exploiting non-linear dynamics of physical substrates for computation. State-of-the-art performance has been demonstrated in both simulation and physical implementation for a variety of proof-of-concept systems exploring novel materials and substrates, including novel nano materials, magnetic materials, neuromorphic hardware and analogue electronics. However, current examples are still limited by components converting between analogue and digital (the real world and the computer), there exists no programming model or methodology to select appropriate substrates and design reservoir computer architectures. In the talk I will motivate my goal to remove any digital pre- and post-processing to exploit the full potential of embodied RC, and outline the vision of a layered, flexible architecture composed of components implementing different RC models – a “reservoir of reservoirs” (RoR) – that can be networked into complex signal processing pathways. I will show the steps that we have taken towards this goal, including a task-agnostic methodology to assess materials for RC performance (CHARC), results for different RC topologies, lessons learnt from different substrates (CNTs, magnets, analogue circuits), and some ventures into applications (NARMA, FPAAs, tensegrity robotics, sensor fusion).