DESI Second-Level AE (SLAE) 640D -> 16D (R42 Phase 2)

SLAE compresses the concatenated 640D ensemble latent (5 BigAE seeds x 128D) into a 16D ultra-rare latent for the BigBounce DESI DR1 anomaly catalog.

Files

  • slae_16d.pt โ€” PyTorch state dict
  • slae_mse.npy โ€” per-spectrum SLAE reconstruction MSE on the 100K OOD set
  • phase2_slae.json โ€” top-10/top-100 indices into the OOD set, val MSE, OOD percentiles
  • phase4_agreement.json โ€” ensemble agreement on top-100 SLAE outliers (cross-validation)

Architecture

Encoder: 640 -> 256 -> 64 -> 16
Decoder: 16 -> 64 -> 256 -> 640

Headline numbers

  • Top-100 SLAE ultra-rare anomalies: 100% agreement with single-seed BigAE p99 thresholds across all 5 seeds (mean fraction = 1.000, std = 0.000)
  • This means SLAE picks the same ultra-rare set that ALL 5 BigAE seeds independently flag

Companion artifacts

  • bamfai/desi-bigae-ensemble-v1 (5-seed BigAE pack)
  • bamfai/desi-spectral-anomaly-detector (single-best BigAE)
  • bamfai/bigbounce-anomaly-catalog (319K anomaly catalog)

Houston Golden, BigBounce program, 2026.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support