Sparse-SNN：浏览器里的类脑脉冲网络 Sparse-SNN: connectome-inspired spiking network in the browser

详细介绍Details

AwareLiquid 的 HF Space 静态演示：在画板上手写数字（MNIST）或选 Fashion 模式，由一个 5% 连接密度、LIF 神经元、掩码取自果蝇连接组统计的稀疏脉冲神经网络在浏览器内识别（ONNX Runtime Web，纯客户端）。页面自述每帧只有约 8% 神经元发放、每次脉冲是加法而非乘加，并给出「约 112× 估算节能」；同页「Honest boundaries」标明是研究产物、与稠密 MLP 有 1.5 / 0.5 分差距、能耗为解析估算非实测。 AwareLiquid’s static HF Space demo: draw a digit (MNIST) or switch to Fashion mode and a sparse spiking neural network — 5% connection density, LIF neurons, masks drawn from Drosophila connectome statistics — classifies it in the browser via ONNX Runtime Web, fully client-side. The page says only ~8% of neurons spike per frame, each spike is an add rather than a multiply-accumulate, and claims a “~112× estimated energy saving”; its own “Honest boundaries” block calls it a research artifact with a 1.5 / 0.5-point gap to a dense MLP and analytic (not measured) energy figures.

适合学什么What to learn

• 游戏/互动 Game / interactive

  可直接玩或操作的教学游戏与 WebGL 演示。 Teaching games and WebGL demos you can play in the browser.

• 其他 Other

  尚未归入上述类型的公开 HTML 页面。 Public HTML pages not yet filed under the types above.

为什么推荐Why we recommend it 公开页 200（2026-09-20）。「5% 密度」「约 8% 发放」「约 112× 节能」均为页面自述，本库未复跑、亦未核实 HF 模型仓与 GitHub 仓库内容，只列为次要来源。页面自己写明是研究产物而非产品，本库照录。视觉上偏工具面板，不进精选。 Public page 200 (2026-09-20). The “5% density”, “~8% firing”, and “~112× energy saving” figures are the page’s own claims — not re-run here; the HF model repo and GitHub repo were not inspected and are listed only as secondary sources. The page itself says research artifact, not production; recorded as such. Visually a tool panel rather than a showpiece, so not featured.