Memristive synaptic plasticity and dopamine-like reward gating in a neuromorphic spiking network for multidimensional taste decoding

Neuromorphic models of gustation commonly focus on sensory encoding and reconstruction, whereas motivational modulation of synaptic plasticity is less often represented explicitly. Here we extend the previously developed TASTE-SNN architecture with a dopamine-like reward-prediction-error (RPE) mechanism that multiplicatively gates the local three-factor update of excitatory Layer-1-to-Layer-2 synapses. Four-dimensional stimuli encode sweetness, sourness, saltiness, and bitterness. A linear critic estimates trial value from Layer-1 activity, a phenomenological hedonic function assigns positive value to sweetness, negative value to sourness and bitterness, and a biphasic contribution to saltiness, and the resulting prediction error drives a decaying trial-scale dopaminergic trace. The gate is permissive: it rescales an eligibility- and error-dependent synaptic update but does not create plasticity in the absence of the underlying local update. In the medium-scale paired experiment, both the unmodulated and +wanting conditions showed decreasing validation loss and maintained Layer-2 firing rates in the 0.06–0.08 spikes/step range. Across 77,839 trials, the dopamine gate had mean 0.972 and median 0.968, ranged from 0.200 to 2.678, and produced nearly balanced cumulative potentiating and depressing update magnitudes. Final off-grid decoding was essentially unchanged ( 𝑅 2 = 0 . 3 7 9 9 for the original network and 0.3792 with wanting modulation; mean L2 reconstruction error 0.3991 and 0.3967, respectively). Mean Layer-2 selectivity index increased from 0.3100 to 0.3626, whereas mean off-diagonal prototype cosine similarity increased from 0.9550 to 0.9591, indicating slightly reduced prototype separation. A label-shuffle control in the +wanting condition reduced 𝑅 2 from 0.3376 to −0.2457. These results indicate that, at the tested network scale and training budget, dopamine-like permissive gating reorganizes aspects of the hidden representation without improving continuous taste-decoding accuracy.