Speaker: Aidan Gomez
Supervisor: Roger Grosse, Mentor: Mengye Ren
Title: Deep Reversible Networks
The primary bottleneck when training extremely deep networks is the limited memory available on accelerated hardware such as GPUs. In this talk, I will discuss work performed with Mengye Ren, Roger Grosse and Raquel Urtasun that addresses this by augmenting the structure of the residual unit to facilitate reversible computation. Using the property of reversibility, we can reduce the activation storage requirement to a constant for arbitrary network depth; enabling massive increase in depth or batch size.
Speaker: Lebo Radebe
Supervisor: Anna Goldenberg, Mentor: Lauren Erdman
Title: Using Random Forests to Predict Paediatric Thyroid Cancer with SickKids Data
Differentiated thyroid carcinoma is extremely rare in children, occurring in less than 1 per 100 000 children less than 10 years old. However, incidence is increasing worldwide. Currently,the only way to diagnose malignancy of tumours is through a Thyroidectomy, a surgery that results in a partial or complete removal of the thyroid gland, located in theneck. Much is still not understood about thyroid carcinoma in general, and pediatric thyroid cancer in particular. According to the current standard of care, only 22-34% of thyroid nodules are malignant; at SickKids, about 46% of patients who undergo Thyroidectomies do not have cancer.
The aim of this project is to improve the understanding and decision making when a clinician decides to operate or perform a biopsy. We are interested in understanding not only which features, but which combination of features are most important in predicting malignancy. Using Machine Learning in a clinical environment, this project addresses issues of interpretability of models, small sample sizes, missing data and using accuracy as a measure of a model.
Speaker: Xi Yan and Caleb Phillips
Eyal de Lara, Mentor: Hossein Mortazavi
Title: Mobile Edge Computing
Current mobile networks are unable to support next generation applications that require low latency, or that produce large volumes of data. Edge computing is a method of optimizing cloud computing system by adding computation and storage capabilities to the edge of the network. This presentation introduces path computing, a generalization of the edge computing vision into a multi-tier cloud architecture that provide storage and computation along a succession of datacenters deployed over the geographic span of the network. We will describe CloudPath, a new framework that implements the path computing vision to optimize access latency and reduce bandwidth consumption, and PathMonitor, a custom monitoring and analytics platform to provide analytics on the new framework.
Speaker: Weidong An
Supervisor: Sven Dickinson, Mentor: Stavros Tsogkas
Title: Improving the Appearance-Based Medical Axis Transform for Natural Images
The medial axis transform (MAT) is a powerful shape abstraction that has been successfully used in shape editing, matching and retrieval. Recently, Tsogkas and Dickinson introduced Appearance-MAT (AMAT) which encodes natural images into medial points by framing the MAT as a weighted geometric set cover problem. The AMAT augments medial points with scale and appearance information, allowing for more informed medial axis grouping, simplification and image reconstruction.
In this project we explore several ways on enhancing the AMAT, such as: (i) run time acceleration; (ii) finding algorithms to maintain cover topology after simplification; (iii) improving the grouping algorithm; (iv) discovering methods to encode texture; (v) applying the AMAT to tasks such as object proposal, image segmentation and scene retrieval.