Deep probabilistic Modelling with Pyro | Chi Nhan Nguyen

Learn how to build robust and reliable deep learning models with probabilistic programming in Pyro, a probabilistic language built on top of PyTorch.

Key takeaways

Neural networks are great, but they have limitations. They perform poorly on out-of-sample examples, are overconfident, and don’t provide good uncertainty estimates.
Uncertainty is important for making informative decisions, and a proper uncertainty assessment of predictions is crucial.
Deep probabilistic modeling involves modeling uncertainties in neural networks.
Probabilistic neural networks provide better calibration of uncertainties and can be used for classification, regression, and sequence prediction.
Pyro is a probabilistic programming language built on top of PyTorch that provides high-level abstractions to sample from many different probability distributions.
Probabilistic models can help obtain robust uncertainties, which is important for critical decisions.
Monte Carlo dropout and variational inference are methods for approximating intractable distributions.
Gaussian processes are a generative model that can be used for regression and classification.
Uncertainty can provide expected variances, which enables planning more efficiently.
Adversarial attacks can be used to test the robustness of neural networks.
Softmax is not a reliable way to model uncertainty in neural networks.
Probabilistic models can be used for time series prediction, sequence prediction, and image classification.
A good model should have a low variance in its predictions.
A confidence interval is a subjective way of representing uncertainty.
Uncertainty can be represented using the KL divergence.
Variational inference is a method for approximating intractable distributions.
The prior distribution represents the prior knowledge of a problem.
Bayes’ Rule is used to update the prior distribution based on new data.
The likelihood function represents the probability of the observed data given the model parameters.
The posterior distribution represents the updated prior distribution after new data is observed.
In probabilistic modeling, we infer probability distributions conditioned on the observed data.
Deep probabilistic modeling can be used for healthcare, biology, and more complex systems.
It’s important to consider shooting from a probability distribution that is closely related to a specific Gaussian processacie
Variational inference can be used to approximate intractable distributions and can be more computationally efficient than Monte Carlo methods.
Autograd is a method for automatic differentiation that is used in variational inference.

Deep probabilistic Modelling with Pyro | Chi Nhan Nguyen

More talks