Memory-Task Alignment in Reservoir Computing

How can we determine whether a computing system remembers the right information for the task it is being asked to perform? Reservoir computing uses the dynamics of a complex system to process time-dependent data. It offers a promising framework for efficient computation at the edge, where data must be processed locally under constraints on power, communications, and computational resources. However, existing measures of reservoir memory and information-processing capacity do not yet provide a general framework for determining whether those capabilities are well matched to the temporal requirements of real-world prediction and decision task.  

This PhD will develop mathematical and statistical methods for memory-task alignment in reservoir computing. The student will investigate how to identify the multiscale linear and nonlinear temporal information required by a task; characterize the corresponding information-processing capabilities of noisy and stochastic reservoirs; and develop statistically principled measures of alignment that account for finite data and system variability.  

A central question will be whether these measures can predict which reservoir architectures or operating regimes are best suited to particular tasks. The methods will be tested using simulated reservoir systems and Dstl-informed defence and security problems, with potential opportunities to connect to ongoing research on novel physical computing platforms (through EPSRC funded project UKRI6477). The project combines applied mathematics, statistics, information theory, dynamical systems, and machine learning.