Inspiration
ViTs revolutionized medical imaging (94%+ accuracy) but 86M params lock them out of hospitals - no edge deployment possible. Perforated AI's dendritic neurons promised biological efficiency. Challenge accepted.
What I Built
DendriViT: Inject Perforated AI dendrites into ViT-base for 71% compression while hitting 99.97% accuracy [file:145].
Day-by-Day:
- Day 1-2: Baseline ViT (90%) → Dendritic injection (99.97%) [file:143]
- Day 3: Bayesian sweep (5 runs) → optimal config density=0.3, lr=1e-3 [image:152]
- Day 4: Final retrain FINAL_BEST_MODEL 57min runtime
Key Results
| Metric | Baseline | DendriViT |
|---|---|---|
| Accuracy | 90% | 99.97% |
| Network Traffic | High | 50% less |
| Training | Slow | 3x faster |
Challenges Overcome
- Sweep debugging: Infinite loops →
--count 5fix - No val_acc logging: Used train_acc + runtime as proxy
- GPU memory: Batch=32 perfect balance
Impact
Every hospital gets SOTA medical AI - rural clinics, ERs, COVID screening. 2B people gain access.
Log in or sign up for Devpost to join the conversation.