Learning Bayesian Statistics BITESIZE | The Why & How of Bayesian Deep Learning, with Vincent Fortuin
Apr 9, 2025
Vincent Fortuin, an AI expert and researcher, dives deep into Bayesian deep learning and how it stacks up against traditional models. He highlights the mathematical nuances and uncertainties in predictions that Bayesian methods bring to the table. The conversation also tackles the lack of cohesive libraries for Bayesian techniques and offers insights on tackling the complexities of real-world applications, particularly in healthcare and climate science. Tune in for fascinating details about improving model interpretability and enhancing AI usability!
AI Snips
Chapters
Transcript
Episode notes
Bayesian Deep Learning Explained
- Bayesian deep learning learns distributions over neural network parameters, yielding uncertainty in predictions.
- It can be seen as Bayesian inference over functions, focusing more on function distribution than parameters themselves.
Deep Learning and Gaussian Processes
- Deep neural networks and Gaussian processes are closely related as both model distributions over functions.
- Bayesian deep learning offers a flexible function distribution beyond the Gaussian assumptions in Gaussian processes.
Choose Bayesian Deep Learning Tools Wisely
- To start with Bayesian deep learning, explore libraries like Laplace, Tihi, and Bayesian Torch for different inference methods.
- Be prepared to try multiple libraries since no unified tool currently dominates the field.