FeaturedDeep Learning · DSCI-D 518Live
AI That Knows When It’s Wrong
A deep-ensemble pipeline for uncertainty quantification. Three WideResNet-16-2 models trained from scratch on CIFAR-10 vote on every image, and the system reports not just a prediction but how much to trust it.
- Total uncertainty measured as predictive entropy of the ensemble mean.
- Epistemic uncertainty isolated as mutual information — the disagreement between members.
- Out-of-distribution testing against CIFAR-100, classes the models have never seen.
- Full stack: PyTorch training and evaluation, FastAPI inference, and a Next.js interface with drag-and-drop upload and ensemble-variance charts.
- PyTorch
- WideResNet-16-2
- Deep Ensembles
- FastAPI
- Next.js
- Recharts
- Docker
- Hugging Face Spaces
> verdict: confident
Concept illustration with example probabilities, computed live in your browser. Real model output is in the live demo.