Explore probabilistic deep learning techniques in TensorFlow through this 29-minute conference talk from ODSC Europe 2019. Delve into the motivations behind probabilistic modeling in deep learning and learn how to apply it using TensorFlow Probability. Discover how to encode expert knowledge into models, support uncertainty in outputs, and fit distributions to neural network weights. Follow along with practical applications and examples, including building regression models, creating chaos models, and implementing probabilistic layers. Gain insights into early stopping, plotting predictions, and inspecting model results. Learn about replacing dense layers with variational alternatives, working with prior distributions, and understanding the "trick" behind probabilistic deep learning. Conclude with a summary of regression using probabilistic layers and an introduction to negative log likelihood in this context.
Probabilistic Deep Learning in TensorFlow - The Why and How - ODSC Europe 2019
Overview
Syllabus
Intro
Overview
TensorFlow dependencies
Building a regression model
Exploring the data
Creating a chaos model
Defining a helper class
Results
Early stopping
Plotting predictions
Takeaways
Replacing Dense Layer
Dense Variational
Prior Distribution
Example Batch
The Trick
Inspecting Layers
Model Results
Model Predictions
Input Predictions
Call Banks
Plot History
Summary
Regression with probabilistic layers
Negative log likelihood
Taught by
Open Data Science
