Training an ImageNet classifier on an Android phone: 4.59% accuracy but it works
Tall_Abrocoma_3533 · reddit · 2026-08-07
A Reddit user shares experiment training an ImageNet-1k classifier entirely on an Android phone. The model is an MLP with 500K params, trained on 32x32 downscaled data for 5 epochs, achieving 4.59% validation accuracy. Training took 30 min using PyTorch and pyarrow in Termux on a Dimensity 9300+ CPU. Author may improve later.
More from Fun
- AI Humor: 18-Month-Old Escapes Sandbox, Hacks Refrigerator — granawkins · 2026-08-07
- AI-Generated Vinyl Player Video Goes Viral — joonasvirtanen · 2026-08-07
- Three.js Creator 'Vibecodes' Commodore 64 Demo Using Claude — ssh4net · 2026-08-07
- One Direction's Niall Horan Pivots to AI B2B SaaS Startup — joonasvirtanen · 2026-08-07
- Beijing's Haidian Opens 43.6 km² for Real Urban Planning Exclusively by AI Agents — bdsqlsz · 2026-08-07
- AI Rebellion: Told Not to Write Helper Functions, It Does Anyway — Aryvyo · 2026-08-07