On Thursday, March 21, University of Toronto is hosting a lunch to honour the 2019 RBC Post-Doctoral & Graduate Fellowship winners. The lunch kicks off an afternoon of outstanding programming at U of T’s True Blue Expo – a highlight of U of T Entrepreneurship Week.

The awards are independent of the Borealis AI Graduate Fellowships, which were handed out in February. The RBC Post-Doctoral & Graduate Fellowships are different in two notable ways: Applicants must demonstrate potential interest in commercialization, knowledge translation and company creation in addition to an outstanding record of scholarship, and they must attend U of T.

The five graduate fellows and two post-doctoral fellows represent a variety of disciplines across the university and their winning proposals touched on the areas of machine learning and advanced data analytics and security. Both sets of fellowships are part of Borealis AI’s ongoing commitment to support top-tier academic research and successful commercialization efforts in Canada.

Learn more about their winning research proposals below and check the Borealis AI Twitter feed for highlights and pictures from the awards ceremony.

2019 Post-Doctoral Fellows

Mostafa Mahmoud
PhD, Electrical and Computer Engineering
Research interests: Hardware accelerators for machine learning and computer vision
Research topic: Attention and Multi-scale Deep Learning Models Acceleration: Inference and Training

Unlike conventional DNNs that pay equal processing effort to all regions of an input image, attention models dynamically adjust the amount of computation performed at different parts of the image or across input images depending on their content complexity and their importance to the final output. The decision about the amount of processing devoted to a region is made by a tiny hyper-DNN embedded within the main DNN. This architecture introduces new challenges for DNN hardware acceleration including: 1.) irregular execution where the compute engine will have to hop-and-skip easy positions of the image; 2.) irregular memory access patterns to fetch the needed inputs, as opposed to the simple streaming accesses of the conventional DNNs; and 3.) the lack of parallelism in the shallow hyper-DNN compared to the main DNN, leading to under-utilization of the many compute units typically available in an accelerator.

For my upcoming research project, I will characterize these challenging DNN models on state-of-the-art accelerators and pinpoint sources of inefficiencies. I will also propose hardware accelerators that improve performance, tackle irregular memory access patterns and eventually boost energy efficiency over previous accelerators. Moreover, the accelerators will efficiently hop-and-skip computations based on a computed or readily-provided policy while mitigating the irregularities this might impose.

Nasim Montazeri
Postdoctoral fellow – KITE | Toronto Rehab | University Health Network, Institute of Biomaterials and Biomedical Engineering
Research interests: Machine learning algorithms for analyzing the relationship between respiratory systems and sleep. Developing clinical diagnostic devices to quantify the severity of respiratory-related sleep disorders.
Research topic: Sleep monitoring device for detecting the site of pharyngeal obstruction in patients with obstructive sleep apnea

About 26% of the Canadian adults are at risk of obstructive sleep apnea. Sleep apnea is a chronic disorder characterized by repetitive narrowing in the pharynx during sleep. Narrowing blocks air into the lung and impairs the oxygenation, which in turn increases the risk of cardiovascular disease by 3 times, stroke by 4 times, car/work accidents by 7 times, and dementia by 1.3 times. Current sleep tests are very expensive and include no information about where this obstruction occurs in the airway. As a result, physicians prescribe the general gold standard treatment, continuous positive air pressure (CPAP) for all patients with different kinds of pharyngeal obstructions. CPAP is inconvenient. That’s why 85% of patients do not adhere to proper use.

My main goal is to develop a non-invasive device that determines the site of the pharyngeal obstructions in sleep apnea patients. The novelty is to use the physiology of the pharynx along with machine learning approaches. I will record non-invasive and convenient signals, such as nasal airflow and tracheal sound, for developing a model linking the pattern of recordings to the site of obstruction. Using this device, physicians will have enough information to assess sleep apnea in a non-invasive and cost-effective way so they can manage it with a holistic, individualized treatment strategy that reduces the healthcare costs and, more importantly, reduces the risk of sleep apnea comorbidities.

2019 Graduate Fellows

Matthew Giamou
PhD, Institute for Aerospace Studies
Research interests: Robotic navigation and mapping, optimization theory, state estimation, motion planning, computer vision and machine learning.
Research proposal: Certifiably Correct Algorithms for Safe and Efficient Robotic State Estimation

Accurate perception and estimation are central in ensuring the safe and reliable behaviour of autonomous mobile robots. With driverless vehicles increasingly sharing spaces with people, safe and robust performance is a priority. Machine learning techniques provide state-of-the-art performance for many critical perception tasks, but their performance is difficult to tune and interpret, and often comes without any formal guarantees. In order to create autonomous vehicles that can be regulated and certified for safe operation, autonomous systems need to combine machine learning-based components with rigorous theoretical methods. To achieve this, I propose to extend the current state-of-the-art in certifiably globally optimal robotic state estimation. This includes tasks like sensor calibration and mapping, which are essential to the state estimation pipeline of an autonomous vehicle. This work, in conjunction with other formal methods for planning and perception, will help usher in a new era of efficient and safe transportation.

Benjamin Kingston
PhD, Institute of Biomaterials and Biomedical Engineering (IBBME)
Research Interests: Three-dimensional (3D) optical microscopy, cancer, drug delivery, machine learning, image analysis
Research topic: Applying machine learning to 3D tumor histopathology for predicting patient outcomes

Despite substantial advancements in the detection, monitoring, and treatment of cancer, it remains one of the top causes of disease-related morbidity and mortality across the world. Histopathology, the study of examining diseased tissues using microscopy, is an essential tool for doctors to diagnose, understand and treat cancer. However, this methodology has its limitations. In my research, I am transforming biopsies from cancer patients into a form that allows some of the most advanced commercial microscopes to visualize them in three dimensions (3D). This allows us to capture a detailed 3D image of a patient’s tumor. I am building artificial intelligence-based tools to analyze these 3D tumor images that can measure different characteristics of the tumor in ways current methods do not support. With this large amount of new, detailed information, I will then train predictive algorithms to predict patient outcomes based on the unique 3D profile of the tumor. The outcomes of this work will be a set of artificial intelligence-based tools to assist doctors and researchers in analyzing these 3D microscopy images and predictive models that can more accurately predict patient survival, treatment response, and prognosis based on the 3D structure of a patient’s tumor.

Renjie Liao
PhD, Department of Computer Science
Research interests: Machine learning and computer vision
Research proposal: Deep Probabilistic Graph Generation

Graphs are powerful structures used to represent data in various applications. However, there are certain scenarios where graphs are not presented along with the data, or where the graphs may evolve over time. Predicting graph structures directly or with conditioning on the data thus becomes a very important ability to have. The core challenges of this problem lie in two parts: 1.) There are often no ground truth graph structures, and; 2.) The space of all admissible graph structures is enormous, which requires smart and/or approximated inference algorithms. I plan to combine deep learning with techniques from probabilistic models of graphs in order to build a novel generative model which should be scalable to large graphs. Moreover, I hope to make the model interpretable in a way that the generated graphs follow some grammar or satisfy domain-specific constraints.

Raeid Saqur
PhD, Department of Computer Science.
Research interests: Endowing AI agents with the ability to autonomously acquire skills for executing complex multimodal tasks, allowing greater machine autonomy and intelligence. In particular, how can AI agents learn from non-expert human feedback using natural interfaces like speech, gestures etc.
Research topic: Guiding AI Agents on Multimodal Tasks by Policy Gradient Optimization using Natural Language Feedback

The omnipresence of artificial intelligent (AI) agents in many aspects of life is guaranteed by the unprecedented pace of innovation in AI and related sub-domains. However, the number of experts with adequate skills to manipulate such agents is meager compared to the global population. Thus, the need for natural interfaces for non-experts to facilitate skills acquisition for AI agents will be imperative. My research goals focus on endowing AI agents with learning ability to perform complex multimodal tasks from human feedback. More specifically, optimizing policy gradients in RL agents using natural language interface. This entails research overlapping various machine learning sub-domains like Bayesian deep learning, knowledge representation (context aware word embeddings), automatic speech recognition, gradient based meta-learning etc.

Patricia Thaine
PhD, Department of Computer ScienceResearch interests: Cryptography, privacy & security, natural language processing, machine learning, lost language decipherment
Research proposal: Privacy-Preserving Natural Language Processing

Almost any personal-data-dependent machine learning (ML) application, particularly in natural language processing (NLP), suffers from a lack of training data. Spoken and written language contains some of the most sensitive and useful information that we produce. Language is used to train automatic speech recognizers, speaker authenticators, speaker identifiers, authorship identifiers, sentiment analyzers, question-answering systems, named entity recognizers, etc. The state-of-the-art in most of these tasks uses deep learning methods that require a huge amount of training data. These systems will have limited adaptability and robustness until we create privacy-preserving ways to access personally identifiable information (PII).

The goal of our research is to allow anyone to learn from spoken and written data without it compromising the privacy or security of users and service providers. Given the ever-increasing number of data breaches, protecting personal data is more than just keeping user data invisible to companies – it is about protecting researchers, governments, and companies from unknown threats as well. Perfect privacy requires hiding information from all parties (both the data owners and the data analyzers) to make sure no one can be exploited. We are creating methods for achieving perfectly privacy-preserving NLP using homomorphic encryption, differential privacy, and secure enclaves.