My toy spiking network completely flunked NARMA-10, but a simple neuroscience trick unlocked a 15x compute bargain. [D]

(Disclaimer: This post was drafted with the help of AI to keep it concise, but the research and work are entirely mine.) I’ve been building a spiking neural network (SNN) engine from scratch on my laptop as a solo project. To see if it was actually tracking anything useful outside of my own custom puzzles, I finally tested it on a standard benchmark: NARMA-10. It was a pretty humbling reality check. It flunked completely. The Failure NARMA-10 requires continuous time-series tracking. When I measured the spiking reservoir, its memory was barely two steps deep when the benchmark needed ten. Tweaking standard dials like input volume or cell lifespans did absolutely nothing to fix it. The Small Fix To get it out of the gutter, I tried a basic concept from neuroscience: heterogeneous wire lengths (adding discrete time delays to the inputs). This spread the past out across the network. It worked well enough to triple the memory depth and finally match a basic line-fitting baseline. It's nothing to brag about yet, but it at least made the network usable on the task. The 15x Efficiency Trade-off I want to be completely transparent—smooth, continuous units still beat this spiking net on absolute accuracy almost everywhere. Spikes are definitely not a magic shortcut to out-compute modern architectures. The only real win is pure efficiency. On a small 512-cell recognition task where the spiking net managed to tie a continuous net in accuracy, I counted the exact internal operations (multiply-and-adds): Standard continuous nets grind through every single cell on every tick, busy or not. My spiking net only does work when a cell actually fires. The rest stays silent. The tally: The spiking net used 15 times less internal arithmetic to get the exact same answer on my standard laptop hardware. Moving Forward This benchmark taught me that spikes don't think better; they just think cheaper when the problem space allows it. Instead of just grinding away trying to force prettier engineering benchmarks or out-accuracy standard models, I'm taking a step back to explore some new, creative avenues for how this engine can actually be utilized.