Uncertainty in deep learning
summary
Builds a Bayesian NN with weight distributions using variational inference (Gaussian mean and sigma, reparameterization trick) so predictions come from many sampled weight sets. Runs on Fashion-MNIST as in‑distribution and Extended‑MNIST as out‑of‑distribution, computes entropy from repeated weight sampling to flag high‑uncertainty or misclassified images, and uses this for active learning decisions. Shows many OOD images get high probability but also higher entropy, notes remaining failure modes and medium‑scale scalability constraints.
AI-written summary · about
chapters (11)
- 0:00Uncertainty in deep learning
- 3:26Brain-inspired neural networks
- 7:47In-distribution vs out-of-distribution
- 10:11Bayesian neural networks
- 13:20Variational inference ELBO
- 19:21Fashion MNIST experiments
- 24:59Evaluating predictions
- 27:07Q&A reparameterization
- 28:10Q&A low-entropy outliers
- 31:34Q&A Bayesian scalability
- 33:21Q&A references
related talks
36mBayesian Deep learning with 10% of the weights
46mKnow What You Don't Know: Tools to Understand Uncertainty in DL
44mAre You Sure about That?! Uncertainty Quantification in AI
05mThe Frontiers of Machine Learning and AI
45mStudying Generalization in Deep Learning via PAC-Bayes
25mDiscovering Your Model's Known Unknowns and Unknown Unknowns
39mBayesian Deep Learning with Edward (and a trick using Dropout)
12mBayesian Neural Networks With Variational Inference in PyTorch