A visual-first map of the paper’s core tension: directly trained SNNs vs conversion-based SNNs, with emphasis on LIF dynamics, surrogate gradients, simulation latency, benchmark scope, and the mismatch between neuromorphic promise and GPU reality.
Games reported
17
Simulation window
64 t
Main comparison
Direct vs Convert
Hardware question
GPU ≠ Loihi
direct training viability
methodological novelty
framing skepticism
The key visual theme: shorter spike simulation windows help, but “human-level control” inherits the burden of a much broader Atari benchmark tradition than a selected 17-game slice.
References
Compact source map for the visuals. Transcript-extracted arXiv IDs: 2201.07211.