Decoder-Aligned Cumulative-Prefix Training for Low-Latency Spiking Neural Network Classification
Low-latency spiking neural network classification depends on the temporal decision state queried by the deployment decoder. For cumulative spike-count decoding, losses attached only to terminal counts, instantaneous spikes, or hidden readouts can optimize states that differ from the cumulative prefixes used for anytime decisions. This paper studies decoder-aligned cumulative-prefix training (DACP) for directly trained SNNs. DACP applies one loss-design rule: supervise the cumulative decoder state used at inference. The resulting curriculum begins with final-horizon count supervision and then adds cross-entropy on cumulative prefix logits in the same space consumed by the anytime decoder. The reported protocol compares Count, Per-step CE, Static Blend, TET, and DACP over five seeds on N-MNIST, CIFAR10-DVS, SHD, and SSC under fixed architectures, optimizer settings, checkpoint rules, and SynOp accounting; robustness checks cover schedules, prefix weights, recurrent depth and width, adaptive recurrent and transformer-style SHD presets, fast-sigmoid substitution, and TET native-readout re-evaluation. On the temporally rich SHD task, DACP raises $\gamma =0.80$ stop accuracy to 64.36%, compared with 47.10% for Static Blend and 36.51% for Count, and raises AATC to 51.84%, compared with 41.72% and 37.56%, respectively. On SSC, DACP gives the highest AATC among the tested losses (27.67% versus 26.64% for Count), but the margin is small. These results support DACP as a decoder-aligned objective for cumulative anytime decoding when class evidence is temporally distributed, while identifying fast-saturating event-image tasks as boundary cases rather than domains of universal DACP dominance.