fly-connectome-chess
An open-source reinforcement learning experiment that uses the actual neural connectome of Drosophila as the topology of an artificial neural network. The model is trained to play chess using a virtual dopamine reward mechanism, exploring the potential application of biological connectomes in machine learning.
Timeline (last 90 days)
A Reddit user (AdilSiddiqui) used the real neural connectome of a fruit fly as a neural network backbone and trained it to play chess using reinforcement learning.
The core design constraint of the project is that signals can only flow along real biological synapses, and the connection matrix is sparse and predetermined by biological structure.
The reinforcement learning mechanism in the project is called virtual dopamine, which applies synaptic weight adjustments only when the model makes a legal chess move or captures an opponent's piece.
The project has been open-sourced to the GitHub repository fly-connectome-chess, providing a training pipeline that can run on a local GPU, with training initiated by clicking the Train button in the Web UI.
The author issued a challenge: if the trained ELO score exceeds his, a PR containing weights.pt can be submitted.
The project does not use a spiking mechanism and still operates within the standard backpropagation/weight update framework, and is therefore distinct from neuromorphic computing at the hardware level.
The project's virtual dopamine triggers weight updates only on legal moves and captures as discrete positive events, with no negative penalties and no explicit modeling of reward prediction error.
The project is a proof-of-concept exploration to examine whether the path of sparse biological topology combined with reinforcement learning is viable, rather than to draw conclusions about performance.
All Facts (8)
The project does not use a spiking mechanism and still operates within the standard backpropagation/weight update framework, and is therefore distinct from neuromorphic computing at the hardware level.
50%UnverifiedThe project's virtual dopamine triggers weight updates only on legal moves and captures as discrete positive events, with no negative penalties and no explicit modeling of reward prediction error.
50%UnverifiedThe project is a proof-of-concept exploration to examine whether the path of sparse biological topology combined with reinforcement learning is viable, rather than to draw conclusions about performance.
50%UnverifiedThe core design constraint of the project is that signals can only flow along real biological synapses, and the connection matrix is sparse and predetermined by biological structure.
50%UnverifiedThe reinforcement learning mechanism in the project is called virtual dopamine, which applies synaptic weight adjustments only when the model makes a legal chess move or captures an opponent's piece.
50%UnverifiedThe author issued a challenge: if the trained ELO score exceeds his, a PR containing weights.pt can be submitted.
50%UnverifiedThe project has been open-sourced to the GitHub repository fly-connectome-chess, providing a training pipeline that can run on a local GPU, with training initiated by clicking the Train button in the Web UI.
50%UnverifiedA Reddit user (AdilSiddiqui) used the real neural connectome of a fruit fly as a neural network backbone and trained it to play chess using reinforcement learning.
50%