This thesis proposes a dual reservoir Liquid State Machine (LSM) architecture with astrocyte control that is bidirectionally coupled and spectrally constrained. The aim is to keep spiking dynamics in a band near the 'edge of chaos' while using only local plasticity inside the reservoirs and readouts. This architecture is evaluated on three datasets (MNIST, N-MNIST, and Fashion-MNIST), aiming to improve accuracy and robustness. The model uses a global estimate of the branching proxy beta^ to push an astrocyte controller that adjusts neuronal thresholds to keep the reservoir activity stable, without the use of any backpropagation through time or the requirement of task specific tuning. On MNIST, test accuracy increases from 96. 3% (single fixed beta) to 96. 9% (single adaptive), 97. 5% (dual uncoupled), and 98. 1% (dual coupled), with expected calibration error (ECE) falling from 4. 2% to 2. 7%. We see similar trends on N-MNIST (96. 1% to 97. 9%), with the increase for Fashion-MNIST being from 84. 9% to 87. 3%, which is comparatively a much harder task. Sweeps (over coupling strength and controller targets) show that accuracy, mutual information, and lag 1 transfer entropy peak near gamma = 0. 10, and that the best performance is near beta* = 1. This model also performs well in robustness tests (such as occlusion, spike noise, etc. ) ; it is consistently more stable than the single or uncoupled baselines. This thesis shows that the proposed dual reservoir architecture, with astrocyte control and bidirectional coupling achieves performance competitive with deep learning models that use much larger feature budgets, all without backpropagation through time, making it ideal for neuromorphic hardware. It also shows how small deep learning components can be combined with spiking neural networks to reap the benefits of both worlds.
Harshankar Gudur (2026) studied this question.