Short Courses

1. Introduction to Graph Neural Networks and Their Generalizations

Instructor(s):
Mustafa Hajij, Assistant Professor of Data Science, University of San Francisco

Graphs are an exceedingly popular type of data for representing relationships between individuals, businesses, proteins, brain regions, telecommunication endpoints, etc. Being a natural object to model relational data, graphs have become an instrumental tool that interface with a vast array of problems ranging from social networks, chemical and physical interactions, drug discovery and road networks. Over the past few years graph-based models, such as Graph Neural Networks (GNNs), have emerged as a set of tools for processing relational data leveraging ideas from classical signal processing, graph theory and recent advances in deep learning. This workshop aims at introducing this topic and provides multiple practical aspects of these models by utilizing a popular open source package PyTorch Geometric. We will start by demonstrating how to build a general GNN with message passing graph neural networks. We then give multiple examples of such constructions and go over construction of convolutional graph neural networks. We then use these constructions to show how to use these networks for practical applications. Finally, we will go over recent advances in GNN and briefly talk about higher order networks.

The learning outcomes for the course are the following:

  • graph neural networks, message passing graph neural networks, and convolutional graph neural networks;

  • constructing a GNN using message passing graph neural networks with PyTorch geometric; and

  • recent frontiers on GNNs research.

2. Bayesian Methods in A/B Testing

Instructor(s):
Nathaniel Stevens, Assistant Professor of Statistics, University of Waterloo

A/B testing has become a ubiquitous method for testing and evaluating product changes in the tech industry. Companies with a culture of experimentation and a “test everything” philosophy commonly run hundreds of experiments per day, engaging millions of users. The statistical framework for A/B testing is consistent with randomized controlled trials: users are randomly assigned to different variants (of a product, ad, promotion, etc.), data is collected on each user, and these observations are used to compare among the variants. Such comparisons are commonly carried out using traditional frequentist methods like hypothesis tests and p-values. However, the comparisons of interest and the available data both lend themselves naturally to Bayesian methods of statistical inference. Accordingly, the use of Bayesian methods in A/B testing is a growing trend and the focus of this short course.

The learning outcomes for the course are the following:

  • the shortcomings of traditional frequentist methods (e.g., misuse and misunderstanding of p-values);

  • the basics of Bayesian inference (e.g., Bayes rule, prior distributions, posterior distributions); and

  • the manner in which Bayesian methods are being used in industry for A/B testing. In particular, how tools like Bayes factors, posterior probabilities, and credible intervals, may be used to inform both the design and analysis of these experiments.

3. Introduction to Deep Learning with PyTorch

Instructor(s):
Carlos Garcia, Data Scientist, University of San Francisco

In this workshop, you’ll learn the basics of deep learning and how to build Neural Networks using PyTorch. First, we will introduce PyTorch tensors and Automatic differentiation package. Then, you’ll get practical experience through coding a fully connected neural network model (FCNN) to solve machine learning tasks (like regression/ classification) using a tabular dataset. We will discuss activation functions and regularization techniques like normalization and dropout. The sessions will use Python 3 and Google Colab.

The learning outcomes for the course are the following:

  • explain and apply their knowledge of Deep Neural Networks;

  • know how to use Python libraries such as PyTorch for Deep Learning applications;

  • build Deep Neural Networks using PyTorch; and

  • train deep learning models.

4. Adapting Pretrained Models for Document Classification  

Instructor(s):
David Guy Brizan, Assistant Professor, University of
San Francisco

More information coming soon!