Digital fruit fly brain learns to read hiragana, hits ~85% accuracy after 100k characters
tkasasagi · x · 2026-09-13
A developer trained a simulated fruit fly brain (FlyWire connectome, mushroom body only) to read hiragana: after 8 hours of training on 100k characters it answers 85% correctly on unseen fonts, and a live web page lets you quiz it in real time.
- Encoding: 16×16 pixels drive 685 olfactory projection neurons with Poisson spikes; 10% of 5,177 Kenyon cells fire, using the same wiring as smell.
- Readout: 96 MBONs split into 46 compartments, one per character; dopamine-based plasticity (Hige 2015, Handler 2019) strengthens correct and weakens chosen compartments on mistakes.
- Curriculum: starts with the 'a' row, unlocks the next only after 75% on 5 held-out fonts; author estimates 70% on all 46 characters.
- The fly 'writes' each answer with its front legs and gets food when correct; the 2-MBON-per-character readout bottleneck explains most remaining error.
Related event: Digital Fly Brain Learns to Read Hiragana with 85% Accuracy(2 posts)→
More from Research
- AIR: Effective-Learning-Rate Scheduler Beats Cosine on 128B-Token MoE Pretraining — stochasticchasm · 2026-10-30
- CMU's Gym-Anything Auto-Builds 200+ Agent Environments from Any Software, Selected by GDP — wellecks · 2026-10-30
- NVIDIA researcher: train robot policies with privileged sim info, improving Pi 0.5 — wellecks · 2026-10-30
- Leveraging Simulator Privileged Information: Three Methods to Improve Robot Models — wellecks · 2026-10-30
- CMU Auto-Builds 25 Robot Simulations with Agents; Frontier Model Solves 0 of 25 Test Tasks — wellecks · 2026-10-30
- DeepMind's f_charton shares his take on the current state of AI and mathematics in Sorbonne talk — f_charton · 2026-10-30