submission.json

{
    "name": "Connectome-JEPA",
    "author": "Eric B.",
    "github_id": "ericdatum",
    "date": "2026-03-30",
    "track": "non_record_16mb",
    "val_bpb": 1.7942,
    "val_bpb_std": 0.0028,
    "val_bpb_seeds": [1.7955, 1.7909, 1.7961],
    "val_loss": 3.0294,
    "artifact_bytes": 1967950,
    "model_bytes": 1920530,
    "code_bytes": 47420,
    "params": 3336051,
    "training_time_seconds": 600,
    "training_steps": 4175,
    "hardware": "8x H100 SXM (Modal)",
    "summary": "Sparse I/O bottleneck architecture inspired by C. elegans connectome (300 neurons, 5084 synapses). 6 sparse hops with gated skip connections, 4-head cross-position attention, JEPA + SigREG training. First JEPA and first biologically-inspired architecture submission. 1.97MB artifact (12.3% of budget). Ablations show sparse I/O bottleneck matters but biological topology does not help for language.",
    "architecture": {
        "type": "connectome-jepa",
        "embed_dim": 768,
        "vocab_size": 1024,
        "num_hops": 6,
        "cross_attn_every": 2,
        "skip_every": 2,
        "head_dim": 128,
        "num_attn_heads": 4,
        "neurons": 300,
        "edges": 5084,
        "sensory_neurons": 88,
        "motor_neurons": 123,
        "connectome_density": 0.0565
    },
    "training": {
        "optimizer": "AdamW",
        "learning_rate": 0.001,
        "lr_schedule": "cosine",
        "weight_decay": 0.1,
        "batch_tokens": 524288,
        "seq_len": 1024,
        "warmup_steps": 100,
        "jepa_weight": 0.5,
        "sigreg_weight": 0.0001,
        "mask_ratio": 0.3
    }
}