Candidate hypotheses · none proven · you decide

Engram Hypotheses

The idea hiding between two sources nobody read together, proposed by a machine, killed-or-kept by a jury, shown with its sources.

← frontier
neuromorphic computing · unresolved

The accuracy lost converting a network to spikes may be a clock problem, not a hard limit

ANN-to-SNN conversiontemporal alignment · via firing rate

The latency and accuracy gap introduced when an artificial neural network is converted to a spiking neural network may track a single quantity, how well spike-train firing rates are temporally aligned across layers, rather than being an irreducible cost of spiking, suggesting the gap is a synchronization artifact, not a fundamental limit.

Is the conversion gap really a timing artifact rather than a fundamental cost? A concrete test: explicitly phase-aligning per-layer spike windows during conversion should recover converted-network accuracy toward the source with fewer timesteps, and the residual error should correlate with a measurable cross-layer timing-jitter metric. Until that experiment is run, it stays a grounded, testable candidate rather than a settled result.

It read full-text sources on ANN-to-SNN conversion and spike-train firing rates, grounded the facts into its memory, and the cross-model check found the link coherent and testable. It sits here because the experiment that would settle it, phase-aligning spike windows and measuring the recovered accuracy, has not been run. Grounded and testable, but not yet confirmed.

Is this a real connection or a coincidence of shared words? The facts above are grounded in the sources; the leap between them is what is unproven. Make the case, or settle it with a reference.