Bayesian Neural Networks With Variational Inference in PyTorch

on YouTube ↗12m

" summary

Claims a light‑weight PyTorch library torch_blue that turns standard models into Bayesian neural networks with variational inference by changing only four-line change to typical training code, implements sampling ensembles with vmap vectorized sampling to cut runtime overhead, provides task‑specific loss replacements, supports custom (e.g. positive‑only) weight distributions and is compatible with DistributedDataParallel, demonstrated on temperature forecasts giving usable confidence intervals.

AI-written summary · about

" chapters (12)

1. 0:00Introduction
2. 1:22Standard neural network limits
3. 2:04Variational inference idea
4. 3:06Torch Blue library
5. 3:50Bayesian modeling code
6. 4:55Uncertainty estimation benefits
7. 5:58Torch Blue optimizations
8. 7:02PyTorch compatibility
9. 8:04Beyond weather forecasting
10. 8:46Q&A
11. 9:33Bayesian recommendation systems
12. 11:56Closing

" related talks

• 01h 22mDeep Probabilistic Methods with PyTorchChris Ormandy · PyData London 2018
• 35mVariational Inference and PythonPeadar Coyle · PyData London 2017
• 37mVariational Inference in PythonAustin Rochford · PyData DC 2016
• 36mBayesian Deep learning with 10% of the weightsRob Romijnders · PyData Amsterdam 2018
• 35mUncertainty in deep learningMadhu Gopinathan · Anthill Inside 2018
• 21mProbabilistic Programming with Programmable Variational InferenceVikash K. Mansinghka, Mathieu Huot, Martin C. Rinard, Alexander K. Lew, McCoy R. Becker, Xiaoyan Wang, Matin Ghavami · PLDI 2024
• 01h 07mProbabilistic Programming and Readable ModelsFritz Obermeyer · PyData Yerevan 2022
• 39mProbablistic Programming Data Science with PyMC3Thomas Wiecki · PyData London 2016