Innovative Projects Realized

Differentially private models for detection of previously seen data

Jumio is constantly facing fraudulent attacks of repeated nature when a series of similar images with minor changes are submitted. To be able to respond effectively it is necessary to be able to learn previously seen fraudulent data. At the same time we have to deal with very sensitive private data and therefore it’s becoming a major concern for multiple reasons.

First, just comparing to all already seen data is not feasible technically due to optimization issues. And second most importantly it would cause legal issues due to privacy concerns and restrictions.

Predicting acoustic and pollutant emissions from combustion equipment using experiments and machine learning

Several engineering equipment ranging from those used for portable power generation, to medium scale gas turbine aircraft engines, and large scale land-based power generation units burn fuel to produce either electric energy or generate propulsive force. This energy conversion takes place inside a combustion chamber which emits noise and combustion pollutants. The objective of the present study is to first perform experiments and analyze data to understand the relations between noise and pollutant emissions. Then, artificial neural networks will be used to perform data mining, developing models that facilitate prediction of both noise and pollution emission from the combustors.

Object occlusion detection

Sometimes an object is partially hidden by a physical object like fingers, book etc. or by digitally pasted artifacts like blurringof an area. The intern will be working on detecting these types of digital and physical occlusions on an object.
At Jumio we encounter documents that are hidden or occluded by such masks. These documents need to be rejected or approved depending on the issuing country. The intern will help us develop a solution which can accurately detect any type of masks and helps us decide if we should approve or reject the masked document based on some decision rules. This project will directly contribute to our world-class product.

Portrait photo segmentation and generation with Deep Neural Network

In this project, the main purpose is to develop a tool for facial image anonymization by replacing the face in an image with a fake, generated face. To have realistic looking generated images, it is essential for the generated face to have a similar pose to the original one and to be seamlessly harmonic to the background. To achieve this goal, the first step is to locate all the faces of humans in images, and then extract key information about the pose (e.i.: eyebrows alignment) to reconstruct a similar face. Finally, a deep network will be trained to draw a fake face with extracted facial information on the location where faces are detected. This can be served as a data anonymization application protecting user’s privacy. In addition, it can be a tool for data augmentation which is a commonly used technique to boost deep neural networks’ performance.

Modelling of Passive Pilot, Pilot Seat and Inceptor for Aircraft-Pilot-Coupling (APC) Induced Oscillation Investigations

Aircraft Pilot Coupling (APC) may arise when airframe structural modes encroach into the frequency range of human senses, biodynamics and control, which is becoming more relevant given the increasingly flexible and optimized airframes of next-generation aircraft. The phenomenon is characterized by oscillations sensed by the aircraft crew and passengers that negatively impact the ride quality. Furthermore, it may impair the pilot ability to perform specific tasks and in some extreme cases might lead to fatalities. The project aims to develop and validate models representative of the relevant pilot biodynamics, aircraft inceptor and pilot seat that will be coupled to a supplied flexible aircraft model with flight control laws. The integrated model will be used to identify signatures of the APC phenomenon and understand the factors leading to its occurrence. The research project will provide Bombardier Aviation with the tool to predict potential APC in an aircraft development program to avoid costly design changes in its later stages, especially during flight testing. It will also allow for implementation and evaluation of solution strategies.

Applying state-of-art NLP models to molecular representation

Molecular generative methods are at the heart of our computational platform. We use cutting edge deep neural networks in order to generate unseen molecules based on existing conditions or molecular space. Since external projects typically require different architectures, we are continuously expanding our generative methods toolbox with new approaches. This project aims to augment our internal toolbox with large pretrained NLP models for molecular representation and generation. During this project, the intern will have to train and benchmark existing architecture with large datasets but also contribute/integrate new models such as the transformers provided by Hugging Face into our existing API.

Acceptation des exosquelettes dans le secteur pharmaceutique : expérience d’implantation d’une nouvelle technologie d’assistance physique pour réduire les risques à la santé au travail

Les troubles musculosquelettiques (TMS) sont une préoccupation importante au Québec et diverses avenues de prévention sont mises de l’avant pour en diminuer leurs impacts : formation du personnel, conception d’aménagements et/ou d’équipements adaptés, ajustement des exigences de production, etc. Dernièrement, les exosquelettes semblent susciter l’engouement pour prévenir ces risques dans les milieux de travail. Cependant, il existe peu de données dans la littérature qui documentent leurs effets – bénéfiques ou néfastes – directement en milieu de travail et/ou à long terme sur la santé des travailleurs. Pour l’heure, il nous est difficile de conseiller les préventionnistes sur la manière d’accompagner les organisations désireuses d’implanter les exosquelettes. Une entreprise du secteur pharmaceutique nous a contacté pour explorer cette avenue de prévention. Ce projet a but pour d’accompagner l’entreprise dans une première implantation d’exosquelettes.

Analyse des processus de communication de Show-Canada

Les Industries Show-Canada inc. ont des problèmes de gestion de projets relativement à la communication. Les divers départements, dont l’ingénierie, le dessin et la production, n’ont pas de communication efficace entre eux. Il n’y a donc aucun suivi des différents échéanciers pour les différents projets. Ceci créé des problèmes au niveau des livrables que la compagnie doit rendre aux clients. Ceci créé également des problèmes au niveau de la gestion des ressources humaines. Ce problème est amplifié par l’incertitude et l’instabilité de la demande. Il y a parfois une surcharge de travail, tandis qu’on peut avoir une pénurie d’ouvrage quelques semaines plus tard. Le but du projet sera d’analyser la situation courante et de proposer des recommandations afin d’améliorer la communication au sein de l’entreprise et, par ce même fait, d’augmenter la productivité.

Making Canadian Homeowners Better Protected Financially from Catastrophic Earthquake Risks

Earthquakes pose significant risks to many Canadian homeowners in British Columbia, Quebec, and Eastern Ontario. Many homeowners in these regions have not purchased earthquake insurance for their homes, and recent surveys suggest that most Quebec homeowners are unaware that earthquake loss is not covered in their general insurance plans. If a devastating earthquake were to occur in one of these regions, homeowners would be left vulnerable, unprotected, and very frustrated. A solution to this current problem is a general earthquake insurance pool. Our objectives include the development of state-of-the-art earthquake loss models that quantify the loss experienced by all stakeholders, should an earthquake event occur. The second loss model will quantify the financial loss to homes in worst-case scenario events. The last objective is to come up with a loss pool that is suitable for all stakeholders. The completion of the previous objectives will be crucial in accomplishing this task.

Effect of carboxylated cellulose nanocrystals on the properties of cement and concrete composites

This project proposes the use of carboxylated cellulose nanocrystals (CNC) – developed and manufactured by Anomera Inc. from Canadian forest – as nano-inclusions to tailor a set of properties in cement and concrete composites towards enhanced performance. Due to its unique properties, CNC can significantly enhance the mechanical performance of cement and concrete composites. The effect of various doses of CNC on the cement hydration, as well as on mechanical properties (e.g., compressive strength, tensile resistance, and flexural capacity) and durability of cement pastes, mortars and concrete will be evaluated in this research. Project results are expected to promote the use of Anomera’s CNC material in construction industry, as well as to promote Canadian forest sector as Canada is considered among the global leaders in the exportation of forest products. Likewise, results will contribute to promoting the development of ecological materials for sustainable construction materials.

Investigation of the light production in liquid Xenon during the first ns

Positron Emission Tomography (PET) is a medical imaging modality enabling the detection of many physiological process, such as the uptake of specific molecule by cancer cells. Inside the patient, 511keV gamma produced by positron annihilation is detected and the annihilated point is mapped. Detecting the interaction time of each gamma-ray with a 10ps timing resolution can dramatically improve the image quality or reduce the amount of radioactive material injected in the patient by reducing the imaging time. Liquid Xenon coupled to Photon to Digital Converter for the detection of scintillation photons is promising solution for reaching 10ps timing resolution. However, it is not clear that enough photons are produced in liquid Xenon during the first 100ps to achieve such a performance. In addition to PET application, measuring light production in liquid Xenon during the first 100ps and even the first ns may enable new background rejection capabilities when searching for dark matter interactions and to a lesser extent neutrino-less double beta decays.

In this project, a small detector called the LoLX (Light only Liquid Xenon) detector is used to study the light produced during first ns in liquid Xenon.

Aeration of hydraulic turbines for increased dissolved oxygen

In warm climates warm temperatures cause thermal stratification in hydropower reservoirs inhibiting mixing and leading to deoxygenation of waters at depth (hypolimnium). Turbines withdrawing water at depth result in low dissolved oxygen (DO) in the downstream flow having a large negative impact on the downstream riverine ecosystem. Legislation in the USA and elsewhere now requires hydropower operators to guarantee meeting minimum DO limits in downstream flows. Andritz Hydro Canada has initiated this project to optimize the elbow deflectors used in draft tube aeration, which is a technological retrofit approach not excessively impacting operation schedules. The main deliverables will be the optimization of the elbow deflectors, through a parametric study of the design parameters involved in maximizing bubble surface area and bubble concentration to result in an increase in dissolved oxygen concentration, and a set of data for validation of Andritz’s Computational Fluid Dynamics model.