Deep Probabilistic Modelling with Pyro

Explore deep probabilistic modeling using Pyro in this 33-minute conference talk by Chi Nhan Nguyen at MLCon. Discover how combining probabilistic paradigms with deep neural architectures leads to more informative results and better decision-making. Learn about the limitations of classical machine learning and deep learning algorithms in modeling uncertainty, and how Pyro, a scalable probabilistic programming language built on PyTorch, addresses these challenges. Gain insights into real-world applications, including time series prediction, multi-sensor systems, and predictive maintenance. Delve into topics such as adversarial attacks, neural network bias, conditional probability, Bayesian networks, Gaussian processes, and variational inference. Follow along with practical examples, including an MNIST implementation, to understand how Pyro handles weight priors, inference, loss, training, and posterior sampling. By the end of this talk, grasp the potential of probabilistic approaches in enhancing machine learning models and their applications in various domains.

Intro
Time Series Prediction
Multi-Sensor Systems
Deep Neural Networks - Limitations
Adversarial Attacks
Neural Networks Predictions
Neural Networks Bias
Conditional Probability
Inference from Data
Probabilistic Regression
Bayes Networks
Gaussian Processes
Probabilistic Neural Networks
Probabilistic Programming Languages
Pyro - Framework
Pyro/Py Torch Example: MNIST
Neural Network Softmax Prediction
Pyro: Weight Priors
Pyro: Inference
Pyro: Variational Inference
Pyro: Loss & Training
Pyro: Sampling from the posterior
Random Noise
Predictive Maintenance Example
Sensor Data 1
Neural Network Prediction
Summary