LANVAR
← The shelf

MouseGPT

Did its thing

Decoding where a mouse is in a maze — from nothing but its neural activity. Machine learning as mind reading, 24 hours on the clock.

✳ 2nd place · PharmaHacks 2024 PythonML 2024
✳ How it actually works
THE MAZE the maze — ● true position, × decoded (lags ~½ s behind) DECODED POSITION LIVE X Y coordinates read straight out of the spike train SPIKE RASTER population spikes → features, scrolling in real time
FIG 1 · the decode, animated — spikes scroll in, the × chases the mouse through the maze
neural recordings spike patterns, mouse in a maze preprocessing features out of spikes decoder model activity → coordinates predicted position where the mouse actually is × the maze (mouse ●, prediction ×) 24 hours on the clock · reading position out of a brain
FIG 2 · the machinery — spikes in, coordinates out
The story

PharmaHacks 2024, Neural Decoding challenge: given recordings of a mouse’s neural activity, predict its position in a maze. With Louis Martinez and Alejandro Naim Monge Rouchdi, the team trained a model to translate spike patterns into coordinates.

It is the purest kind of ML problem — a signal that obviously contains the answer, and no obvious way to get it out. Twenty-four hours of feature engineering, model tuning, and validation later, MouseGPT took second place.

Next on the shelf Coal to Goal →