Reservoir Computing for Efficient Edge Computing

Artificial Intelligence (AI) and Machine learning (ML) have been recently employed in real-world applications including consumer electronics, self-driving cars, building or traffic management systems, environment monitoring etc. These algorithms are capable of learning, modelling and processing nonlinear relationships between inputs and outputs in parallel. Technologies that use AI or ML provide the ability to sense, react and learn from the external environment, enabling systems to be more flexible, dependable and more efficient.

Current implementations of such systems, even when based on novel neuromorphic computing paradigms, are based on classical digital computing making them power hungry, fragile, and hard to interface to the real world. Reservoir computing (RC) can help overcome these issues, particularly by being able to perform embodied computation that can directly exploit the natural dynamics of the substrate, thereby dramatically reducing power requirements and providing a natural fit to certain computational tasks. Recently, RC has proven to be an efficient approach for signal processing and dynamical pattern recognition. State-of-the-art performance has been demonstrated in both simulation and physical implementation.

This research project aims to create a flexible architecture based on RC theory previously developed in the research group to support efficient, flexible and feasible multi-reservoir computers. Feasibility and performance of this system will be demonstrated on a range of applications, e.g., starting from a smart microphone implementing noise cancelling, sound source isolation and speaker identification, and building up to a universal 'edge processor' that can process data from diverse sensors, trainable for different tasks and contexts.

Goals:

- To apply the reservoir computing (RC) concept to create efficient hardware for sensor data processing at the edge.

- To develop algorithms and methodology for processing and performing inference using sound input.

- To create hardware architecture to demonstrate the RC-on-Edge concept in a real-world application.

The research group has a strong track record in RC applied to a range of substrates, including digital computers, analogue hardware, and nanomaterials. This PhD will build upon this and push the boundaries towards efficient algorithms and implementations of RC. The methodology will be to focus on the application area of edge processing and inference from sensors recording audio signals. The initial focus will be to survey the literature and evaluate the state-of-the-art of pre-processing methods combined with machine learning approaches. One anticipated gap in the literature is that it is often at the complex and complicated pre-processing stage where most problems are solved, and there is no real requirement for the additional, equally complex machine learning stage. Our approach will be to investigate whether pre-processing can also be achieved with RC in a more flexible and efficient manner. The next steps will then be identified based on the outcomes of this stage, according to the objectives stated.

Benefits extended to UK industry from this PhD project will result from contributions to self-aware reconfigurable fault-tolerant systems and enable commercial applications of resilient (safety critical embedded), connected (low-power IoT) and energy efficient (exploiting dark-silicon) systems. It will contribute to the latest EPSRC cross-ICT priorities: future intelligent technologies (evolvable hardware), safe and secure ICT (resilient systems that can deal with ambiguous data, transient faults and uncertain environments) and cross-disciplinarity and co-creation working at a cross-section of electronic engineering, neuromorphic hardware and biology.

Collaboration with the University of Sheffield on the MARCH project (EP/V006029/1).