Scientific-Mood FAIR Challenge - ML Challenge 2025 (Year 2)
Model weights for all three tracks of the 2025 NSF HDR Scientific-Mood FAIR ML Challenge.
📦 Code Repository: GitHub
Track 1: A3D3 Neural Forecasting
Task: Predict future neural activity (10 timesteps) from historical recordings across two non-human primate subjects (Affogato & Beignet).
Final Score: Total MSE = 38,096 (improved from baseline 50,257)
Models
| Architecture | Description |
|---|---|
| TCNForecasterLarge | Temporal Convolutional Network with CausalConv1d, cross-attention, channel embedding |
| TransformerForecasterPreLN | Pre-LayerNorm Transformer, 8 heads, learnable positional encoding |
| NHiTSForecaster | N-HiTS with MaxPool downsampling, multi-scale MLP stacks |
Key Techniques
- CORAL Domain Adaptation for OOD generalization
- Multi-seed ensembling with per-domain routing
- Domain-specific post-processing (Gaussian smoothing)
Weights
a3d3/neural_forecasting/model_weights/affi/— 9 models + normalizationa3d3/neural_forecasting/model_weights/beignet_public/— 11 models + normalizationa3d3/neural_forecasting/model_weights/beignet_private/— 6 models + normalization
Track 2: iHARP Predicting Coastal Flooding Events
Task: Predict coastal flooding events from hourly meteorological/tidal data.
Model
- Hybrid ensemble: F1-Push Ranker (0.6) + XGBoost G7v2 (0.4) for per-window; G7v2-only for per-day
Weights
iharp/predicting_coastal_flooding_events/final_submission/hybrid_final_v1/— XGBoost booster, F1-Push Ranker bundles, scaler stats
Track 3: Imageomics Informatics — Beetles Scientific
Task: Species classification with out-of-distribution generalization using BioCLIP backbone.
Overall RMS (CRPS): 0.5866
Model
- MLP LP-FT (V124c): Linear Probing then Fine-Tuning on BioCLIP, no bias correction
Weights
imageomics_informatics/beetles_scientific/submission/mlp_lpft_fold1_fp16.pt— TorchScript JIT, FP16 (817 MB)
Requirements
- Python >= 3.10
- PyTorch >= 2.0.0
- NumPy >= 1.24.0
- SciPy >= 1.10.0
License
MIT
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support