We present a Wilson–Cowan reservoir computer (WC–RC) that treats a retinotopic excitatory–inhibitory neural field as a structured reservoir. Travelling waves and bounded oscillations provide an interpretable spatiotemporal basis, while a two-stage sampler (40 sites \(\times\) 200 steps) exports 8000 features per input—a \(>99\%\) reduction with unchanged integration cost. On MNIST and Fashion-MNIST, recurrent readouts trained on these features achieve strong performance within this fixed export budget: Att-LSTM reaches \(85.4\%/79.0\%\), while GRU and vanilla LSTM yield similar accuracies (all with tight Wilson 95% confidence intervals). A simple MLP readout performs markedly worse (\(45.6\%/54.2\%\)), whereas a compact ridge classifier still attains non-trivial performance (\(71.2\%/72.7\%\)), indicating that the exported WC–RC representation is partially linearly decodable but benefits further from temporal modelling. Selective suppression of lateral couplings (\(\lambda \in \{0,0.25,0.5,0.75,1\}\)) shows that reinstating wave dynamics improves recurrent models while degrading the MLP, supporting a functional role for propagating dynamics. A complementary neighbour-coupling ablation, implemented via a diffusion-like scaling factor \(\psi\), shows that increasing lateral spread beyond the baseline regime reduces late-time spatial variance and degrades recurrent-readout accuracy, indicating that useful computation depends on balanced wave dynamics rather than maximal smoothing. At matched exported-feature and readout budgets, ESN baselines underperform (\(\approx 64\%/\approx 73\%\)), whereas compact CNNs achieve higher accuracy but with larger parameter and MAC budgets and without wave interpretability. Supplementary analyses confirm numerical fidelity and relate performance gains to propagation coherence. We outline a fixed-point streaming-convolution mapping for FPGA/ASIC deployment, positioning WC–RC as an interpretable, energy-aware reservoir for neuromorphic vision.
