We are working with researchers at IREQ to accurately predict the global horizontal irradiance (GHI) in Quebec and the north-eastern US from zero to six hours in the future, using images from a geostationary environmental satellite. The GHI is the total amount of radiation received from above by a horizontal surface per unit of time. Better predicting the GHI will primarily improve our understanding of the potential for electrical production using solar energy and provide additional tools to manage the electrical grid. One of the main challenges in this project is the scarcity of gold truth GHI signals, as the number of pyranometers in the territory of interest is extremely small.
We collaborated with CDPQ to explore machine learning-based investment strategies for predicting allocations of a portfolio of stocks over a period of time, based solely on the past prices of those assets, with the goal of maximizing the Sharpe ratio. Those strategies, referred to as momentum-based, showed interesting results in the recent literature and we wanted to verify their applicability to the investing context of CDPQ. We trained models for predicting allocations of S&P 500 stocks, taking into account the investment constraints of CDPQ. While those models outperformed the S&P 500 index, their performance was mostly at par with that of a uniform stock allocation. We proposed ideas based on continual learning to try to overcome the observed limitations of those models.
Dialogue provides a virtual healthcare platform that aims to assign a patient to the right practitioner thanks to a chatbot whose goal is to collect all relevant information about a patient’s symptoms and antecedents. The system used by Dialogue is ruled-based and is therefore difficult to extend to include new pathologies or new medical knowledge. It also asks a lot of questions and doesn’t always collect all relevant information. Mila built a new model, based on reinforcement learning, that collects much more relevant information while asking a smaller number of questions. The training of this model includes a reward shaping function whose goal is to mimic the reasoning of a senior physician.
American Family Insurance is interested in using deep learning models to assist its insurance professionals by accelerating the completion and the verification of claims for its automotive insurance line of products. Those models need to handle multimodal inputs such as text, categorical data, and images. The models will be designed to predict or verify the values of attributes describing individual vehicles involved in an accident, as well as overall claim-level attributes describing properties of the accident.
Optina Diagnostics is developing the Retinal Deep Phenotyping Plateform™ for detecting the cerebral amyloid status and other key biomarkers for Alzheimer’s disease (AD). Given that the eye is directly attached to the brain, it can provide insights into neurological pathologies. The goal of the project is to use Optina Diagnostics’ retinal hyperspectral images to detect signs on the retina of the presence of amyloid beta in the brain. This approach is less invasive and less costly than brain positron emission tomography (PET) scans. Retinal scans thus offer the possibility to improve the availability of tests that could lead to earlier AD diagnosis, and therefore improve the quality of life of impacted patients. We are using the results of a clinical study of a few hundred patients for whom both brain PET scans and Optina Diagnostics’ retinal hyperspectral images are available. Physicians analyzed the PET scans to categorize the amyloid beta status of the patients. This project has many challenges, including the limited number of patients, the large size of the retinal hyperspectral cubes (in terms of spatial and spectral resolution), and the uncertainty of the way amyloid beta manifests itself in the retina. We are collaborating with Optina Diagnostics’ machine learning team to accelerate the development of deep learning models, and ensure that this team can maintain, evolve and extend the results of the project.