Spike Pocket trains a leaky-integrate-and-fire neural classifier with straight-through surrogate gradients and Poisson rate coding. A conventional dense network with exactly the same weights and biases is the matched control.

Evaluation includes clean accuracy, Gaussian-noise accuracy, hidden spike rate, and a synaptic-activity proxy. Spike activity is not presented as measured hardware power; actual energy depends on a neuromorphic deployment target.

Both models contain exactly 4,810 parameters. The LIF network reached 98.15% clean accuracy versus 97.41% for the dense control; both reached 90.74% under Gaussian noise at standard deviation 0.20. Hidden neurons fired on 41.40% of evaluated timestep opportunities.