Impact Report 2020-21

A team led by Professor Reihaneh Rabbany, Canada CIFAR AI Chair and Assistant Professor at McGill University, along with experts in cybercrime and criminology, recently launched Infrared. This project aims to sift through the online advertising market to find anomalous organized activities characteristic of human trafficking rings involving multiple potential victims. In line with Mila’s mission for the responsible development of AI, the research team is acutely aware of the ethical considerations involved and is grounded in victim-centric governance principles. View the news coverage (in French)

Biasly AI deploys a natural language processing algorithm onto online text to identify not only racial or gender bias but also debias problematic sentences. Advances in natural language processing will contribute to future research on the automatic detection of bias in written texts and may eventually allow these algorithms to analyze all forms of texts.

In spring 2020, Mila and UNESCO launched a call for proposals to identify blind spots in AI policy and program development. The collective work, to be published later this year, will explore creative, novel and far-reaching approaches to provide world leaders, policy-makers and civil society members with perspectives that will be critical to ensure the development of AI reaches its full potential in accordance with ethical values and human rights.

All over the planet, companies and their subcontractors exploit men, women and children to reduce the costs of the goods we consume daily. So, how can we detect whether supply chains of goods produced with components from multiple regions of the world comply with international labor standards? In collaboration with The Future Society, which is supported by the international human rights group Walk Free, researchers from Mila are currently developing an algorithm capable of performing compliance analysis of reports issued by companies regarding the use of modern slavery in their supply chains. These reports are required in some nations by modern slavery legislation. This algorithm would make it possible to compare these reports with the standards in force, thus contributing to the increased accountability of companies in this area.

Five projects co-led by Mila scientists are being supported by the CIFAR AI & COVID-19 Catalyst Grants program:

• PanXcea: PANdemic Prediction with X-ray-based COVID-19 External Analysis Marzyeh Ghassemi (Vector, UofT), Joseph Paul Cohen (Mila, UdeM), Chris McIntosh (UofT)
• Leveraging Biomedical Knowledge Graphs for COVID-19 Drug Repurposing Strategies Jian Tang (Mila, HEC), William L. Hamilton (Mila, McGill), Yoshua Bengio (Mila, UdeM), Guy Wolf (Mila, UdeM), Yue Li (Mila, McGill)
• Modeling the Transmission of SARS-CoV-2 Between Zoonotic Sources on a Gene Level Guillaume Rabusseau (Mila, UdeM), Vladimir Makarenkov (UQAM), Bogdan Mazoure (Mila, McGill)
• Machine Learning Against COVID-19: Accelerating Small Molecule Drug Discovery Sarath Chandar (Mila, Polytechnique Montréal), Matthew Taylor (Amii, UofA), Sai Krishna (99andBeyond)
• Planning as Inference in Epidemiological Dynamics Models Frank Wood (Mila, UBC), Benjamin Bloem-Reddy, Alexandre Bouchard et Trevor Campbell (UBC)

Ioannis Mitliagkas, Mila Core Academic Member and Assistant Professor at DIRO of UdeM received Mitacs grants for two COVID-19 related projects:

• Using machine learning to improve treatment against COVID-19;
• Novel molecular representation, molecular generation and data mining strategies for SARS-CoV-2 treatment design.

Professor Bengio was also awarded by the Canada First Research Excellence Fund ‒ IVADO for another COVID-19 project:

• Integrating genomics and AI to accelerate drug discovery against COVID-19

A recent study led by Professor Reihaneh Rabbany and her students proposes a modified version of the standard SEIR model. Taking COVID-19 infected flights to and from Canada into account could enable early detection of outbreaks, more accurately estimate the reproduction number of the disease and better evaluate the impact of travel restrictions and the implications of lifting these measures.

Social distancing measures induced by the COVID-19 pandemic led Professor Danilo Bzdok and his colleagues to seek to understand the effects of loneliness and lack of social support on our brains and overall health. Professor Bzdok's neuroscience research has resulted in two distinguished publications in leading scientific journals, Nature Communications and Cerebral Cortex. The first study reveals a neural “signature” in the brains of lonely people that make them distinct in fundamental ways, based on variations in the volume of different brain regions and on how those regions communicate with one another. The second study found that the brains of individuals lacking social support have distinct neural substrates that make them more vulnerable to loneliness, addiction and stress. This work underscores the critical need to deepen our understanding of the consequences of social isolation on mental and physical health.

Optina is developing a platform for the early detection of Alzheimer’s disease using retinal imaging. With retinal images provided by Optina, Mila scientists are working with the Optina team to develop deep learning models to characterize fundus biomarkers with more accuracy and accelerate early detection of this disease.

D-BOX is redefining entertainment experiences with its immersive haptic technology. The Quebec-based company sought the expertise of our applied research scientists to better understand how machine learning can be used to process music. D-BOX's goal is to accelerate the process of designing haptic cues for a piece of music. Working as consultants, Mila team members provide recommendations to D-BOX on preferred approaches to building a proof of concept for music annotation.

Mila’s applied research scientists are working closely with researchers at IREQ to more accurately predict solar irradiance over a given territory, which is the amount of energy the sun releases when it reaches the ground. The goal is to obtain forecasts that are better than those of weather models over a period of zero to six hours. Such an improvement will deepen our understanding of the potential for electrical production using solar energy and provide tools to manage the electrical grid, given that its production fluctuates greatly over time.