Machine Learning

(You can also browse all our machine learning materials here.)


“A map of current machine learning frameworks”

Webinar (2023-Oct-03) by Marie-Hélène Burle

We are in a period of active development of new deep learning techniques, adding to the already mature area of traditional machine learning. This is leading to a vast and ever evolving field of implementations which can be disorienting. This webinar guides you through a map of the current frameworks, organizing them based on their domain (machine learning vs deep learning) and the languages required to use them. We also talk about the various automatic differentiation options available. To narrow such a large topic, we are limiting the map to frameworks that can be used from Python, Julia, and R.

“TensorBoard for objection detection models”

Webinar (2022-Oct-12) by Jillian Anderson

TensorFlow remains one of the most popular machine learning frameworks for researchers from across disciplines. The TensorFlow ecosystem’s size & flexibility makes it a powerful solution for tasks such as machine translation, image classification, and voice recognition. However, the same things that make the TensorFlow ecosystem powerful can also make it intimidating for new users. In this webinar we focus on one part of the TensorFlow ecosystem - TensorBoard. TensorBoard is a toolkit that allows users to monitor training and evaluate trained models through a visual interface. In other words, with TensorBoard you can explore your models and their performance by looking at graphs and images rather than by combing through logs. In this webinar, I introduce how TensorBoard can be used on the Digital Research Alliance’s HPC clusters to (1) monitor a model during training and (2) evaluate and compare models once training is complete. This webinar focuses on the TensorBoard tools specifically relevant to object detection models, with examples from a project in ecological monitoring that utilized the Cedar cluster for training. However, the general concept of using TensorBoard for model monitoring is applicable to projects from across disciplines and HPC clusters.

“Everything you wanted to know (and more) about PyTorch tensors”

Webinar (2022-Jan-19) by Marie-Hélène Burle

Python already has several multidimensional array structures – the most popular of which being NumPy’s ndarray – but the particularities of deep learning call for special characteristics: the ability to run operations on GPUs and/or in a distributed fashion, as well as the ability to keep track of computation graphs for automatic differentiation. PyTorch tensors provide these and much more, can be easily converted to/from NumPy’s ndarray and integrate well with other Python libraries such as Pandas.

“Upscaling with PyTorch”

Webinar (2021-Nov-24) by Marie-Hélène Burle

Super-resolution (the process of recreating high-resolution images from low-resolution ones) is an old field, but deep neural networks have seen a sudden surge of new and very impressive methods over the past 10 years, from SRCNN to SRGAN to Transformers. This webinar provides a quick overview of these methods and shows how the latest state-of-the-art model — SwinIR — performs on a few test images using PyTorch as our framework.

“Introduction to deep learning with fastai”

Webinar (2021-Apr-14) by Marie-Hélène Burle

fastai is a deep learning library with high-level components that can quickly and easily provide state-of-the-art results in standard deep learning domains. It also provides researchers with low-level components that can be mixed and matched to build new approaches.

fastai aims to do both things without substantial compromises in ease of use, flexibility, or performance. This is possible thanks to a carefully layered architecture, which expresses common underlying patterns of many deep learning and data processing techniques in terms of decoupled abstractions.

This webinar takes a closer look at the features and functionality of fastai.

“Machine learning in Julia with Flux”

Webinar (2020-May-13) by Marie-Hélène Burle