Skip to content
flycade

Inside the fly

A fly connectome. A learned controller.

Game state

Health, distance, actions

Fly connectome

MaleCNS

Game controls

Move, defend, attack
Neural nodes
166,700
Input values
14
Available actions
6

A simplified model, trained on game-state values. The public page replays its recorded decisions.

What this recording shows

A saved policy plays one easy Street Fighter II Turbo round: Ryu against Zangief, on SNES, Turbo 3. The 216 recorded decisions include the game image, consumed controls, stereo audio and sampled neuron states. Playback follows game time; no model runs or learns in this public page.

The model

The policy uses a fixed MaleCNS connectome with 166700 nodes. It receives 14 game-state and previous-action values and chooses among 6 actions held for 4 game frames. Learned input interfaces, shared internal gains and memory, and an action decoder were trained by imitating a 361-decision scripted demonstration for 6 epochs.

This is a simplified recurrent rate model built on anatomical connectivity. It does not reproduce every biological process, and it does not see the game through the fly’s eyes.

Reading the colours

The 12,000 visible points are sampled neurons, not individual synapses. Cell-family colours come from broad anatomical annotations: optic lobe, visual relays, central brain, nerve cord and motor groups, and brain–cord links. They do not identify Street Fighter input channels.

Brightness follows the magnitude of the measured state. A short halo follows changes between consecutive measurements. The Activity mode uses blue for a negative state and green for a positive state; these colours do not classify neurons as inhibitory or excitatory. This capture has roughly 15 measurements per second at game speed. The display interpolates between updates with a 20 ms time constant and lets the halo decay with a 65 ms time constant.

What is established

Across 4 nearby starts of this same round, the connectome policy went from 0/4 wins before imitation to 4/4 after learning, then 4/4 after reloading. These starts were inspected during development. This page shows one of them: 1/1 recorded round victory. A GRU reference reached 3/4 on the same 4 starts.

This small experiment does not establish statistical superiority, general gameplay ability, a full match victory or game completion. Difficulty, character and opponent generalization remain unmeasured.

Follow her progress →

Provenance and files

Checkpoint epoch-0006 from experiment a9c832f41f264d7ca3e4d0ab7886773d. Capture 3458c60b712b4413bd75b2cb7b1e5817.

Sampled anatomy · 1.1 MB · Recorded telemetry · 16.8 MB. Files are compressed in transit and contain no model weights or ROM.

Google Research’s MaleCNS announcement describes the anatomical dataset underlying this experiment.