AntennalLobe_Pong

AntennalLobe_Pong v1

Synaptic weights for the Pong game of BeatTheFly -- A Smart Fruit Fly is playing Pong against you: a spiking network wired as the real Drosophila antennal lobe connectome that moves a Pong paddle in real time.

The anatomical connectome gives you wiring, not synaptic strengths. Ours are trained.

Synaptic weights trained with PHCSSM parallel-scan mode, deployment in sequential RSNN mode (PHCSSM).

made by Po-Han Chiang @ NYCU

What the antennal lobe does in the fly

The antennal lobe is the fly's first olfactory relay: it separates and sharpens smells. Its gate is octopaminergic — 2 cells over 703 real synapses, as the antennal lobe has no dopaminergic neurons — and what they write into the fast weight fades within about two frames. In the real fly, the antennal lobe is the first stop for smell. It sorts signals from the olfactory receptor neurons into glomeruli (one channel per odor-receptor type) and passes them on to the mushroom body and lateral horn. It filters sensory input rather than making decisions.

Architecture

• Wiring: MaleCNS v1.0 antennal lobe -- 3,874 neurons (Receptor neurons (ORN, TRN, HRN) 2,730; Projection neurons 700; Local and interneurons 444) and 465,373 neuron-to-neuron connections. The connectivity mask is fixed to the connectome; 394,286 connections carry a nonzero weight and 0 weights lie off the connectome.
• Dale's law: one sign per presynaptic neuron from predicted neurotransmitters (excitatory 3,355, inhibitory 458, modulatory 61); 0 weights violate it.
• Input: each frame provides 6 numbers seen from the fly's side of the court: the ball's position and velocity and the positions of both paddles. A linear encoder with LayerNorm drives only the input population: 2,730 ORN/TRN/HRN receptor neurons.
• Neurons: leaky integrate-and-fire with per-neuron leak, threshold and reset; synaptic delay of one step.
• Readout: linear map from the membrane voltage of the output population only, 700 ALPN+ALON projection neurons, to 3 paddle commands (stay, up, down).
• Neuromodulatory gate: 2 of this region's own known-neuromodulator cells (octopamine) gate the current into the output population, masked to the 703 real gate-cell -> output synapses in the connectome.
• Fast weight: a value written onto the 38,295 real input -> output synapses, gated by the same cells, read back into the output neurons' voltages. It is read with the current frame's input spikes and written with the same frame's, and it decays: one trainable factor per output neuron, a decay time constant of about 1.82 frames (30.4 ms at 60 fps).
• Deployment: sequential RSNN mode, one timestep per frame at 60 frames per second, with the neuron state carried across the whole game.

Data sources

Data source: actions of a scripted Pong player.

Evaluation

20 games to 11 points against each scripted player, the fly playing its top command every frame:

On 131,072 held-out frames the fly's choice matches the scripted player's action on 97.4% in sequential RSNN mode, and the parallel-scan and sequential modes choose the same action on 99.92% of frames.

Files

• manifest.json -- every tensor (file, dtype, shape, bytes), the model scalars and a connectome audit.
• info.json -- neuron metadata used by the page (cell classes, hemispheres, soma coordinates).
• selfcheck_<precision>.json -- reference observations and logits that the page replays when it loads.
• fp16/, fp32/ -- raw little-endian arrays.

Two precisions are listed in the manifest: fp16w32 (default, 2.9 MB: float16 for the readout matrix dec_w, float32 for the recurrent weights and all other tensors) and fp16 (2.1 MB, recurrent weights in float16 as well).

The recurrent weight matrix W[dst, src] is stored in CSC order by source neuron (W_colptr, W_rowidx, W_vals): each step multiplies W by a sparse binary spike vector, so the engine visits only the columns of the neurons that spiked. in_idx lists the input population and out_idx the output population. enc_obs_T [6, H] is the observation encoder and dec_w [3, 700] is read against the output population's voltage.

Numerical check: 0 of 35,826,752 spike bits differ from the reference on the same weights (2,048 held-out frames and a 7,200-frame closed-loop game).

Limitations

The fly learned by copying a scripted player: it gets no reward and does not plan ahead. It has been tested only in this simulator and only against scripted players, and it loses every game to a perfect one.

License and attribution

Weights: CC-BY-NC-4.0. They are derived from the MaleCNS v1.0 connectome (Janelia FlyEM and collaborators, https://male-cns.janelia.org/, CC-BY-4.0) and trained with PHCSSM (https://arxiv.org/abs/2604.01295); please credit both.

Citation

PHCSSM: https://arxiv.org/abs/2604.01295