Anomaly detection from system logs through deep learning

During the last decade, we observe in organizations a surge of numbers of cyber-attacks originating internally. In this project, we aim to develop deep learning models to detect suspicious activity (such as malicious events, system failure or attacks) from log data generated by the Desjardins ecosystem.

Adapting Ship Refit and Maintenance Planning Optimization to User Preferences

Thales Canada is developing Refit Optimizer, a prototype solution for managing military ship refit projects and maintenance operations. In particular, this tool offers a scheduling optimization component, which decides when each task needs to be executed. During its development, three different criteria were identified according to the challenges faced by the shipyards: minimization of the project's duration, minimization of the costs associated with overtime labor, maximization of the schedule's robustness.

Technical and Procedural Knowledge Extraction with Question Answering

Engineering organisations like Thales rely on large quantities of technical knowledge. When resolving a technical problem, for example, users have to follow a multi-step procedure in which the steps are described with various levels of detail, may not be up to date, or may not target the exact problem they are facing. In this context, our project aims at developing AI language and knowledge representation models that will represent technical knowledge from an unstructured dataset of procedures, in order to identify missing knowledge through a question-answering process.

Feature and Subgraph based Graph Neural Network (GNN) Explanations

This project will develop a graph neural network model that generates a prediction (e.g. detects anomalies) together with a set of explanations as to what the model based its prediction on. The model applies to dynamic graphs (such as wireless communication networks), and the explanations generated are either based on the properties of the graph nodes (e.g. parameters of network cells, such as traffic or power consumption) or the connectivity structure of the node (e.g. the state of the neighboring cells).

Consolidating High-Frequency and Textual Data for Optimal Anomaly Detection in Derivative Markets

In the last few years, a high increase in the interest of traders and investors towards financial instruments directly led to an important augmentation of the information received daily by exchanges. Exchange regulators, who constantly monitor markets to unveil potential infractions, traditionally perform their investigation manually and the notable growth in market activity represents an important risk of fraudulent events going unnoticed. In response to that new reality, exchanges around the globe are establishing automated surveillance systems that track market activity.

Entraînement d’un agent conversationnel en assurance

Koïos Intelligence est à la source de l’agent conversationnel Olivo qui vise à offrir une expérience interactive, guidant l’utilisateur au travers des processus de prévente, de vente et d’après-vente pour tous les types d’assurances. Bien que l’outil soit dans un état avancé tant au niveau de la conversation écrite qu’orale, et ce aussi bien en français qu’en anglais, son amélioration se heurte aux exigences computationnelles lourdes pour l’entraînement des modèles d’apprentissage automatique sous-jacent.

Automated Open World Gameplay Exploration with Curiosity-driven Reinforcement Learning

Computer games are one of the main use cases for AMD graphics cards. To ensure the highest performance and quality possible, AMD has to test its graphics hardware and software on dozens of game titles and hundreds of system configurations. While some game titles provide built-in automated benchmarks, the majority of gameplay testing is manual and results in significant effort expended.Open world games, in which players have the freedom to explore a large and expansive map in a non-linear fashion, have seen a growth in prominence in the gaming industry in recent years.

Reinforcement Learning based Graph Convolutional Recommender Systems

This project aims to use and experiment deep learning technique on modern recommender systems such as Graph Convolutional Network. The purpose of this implementation will be to drastically improve recommendation structure’s benchmark. This will allow extract user’s embedding by mapping from pre-existing features that describe the user such as ID and relevant attributes.In this project students will be integrated as a member of the advanced analytics research team that includes multiple PhD holders in relevant domains.Students would work on the following main topics: 1.

Improving the First Pass Yield of an industrial electroplating line through a combined Design of Experiments and Causal Inference integrated with Deep Learning algorithms (DLs)

The performance of an industrial electroplating line is evaluated using the First Pass Yield (FPY).Improving the FPY of an industrial electroplating line is complex and depends on lots of environmentaland technical parameters. These parameters have different nature that makes it hard to assess theinteraction between them and consequently to detect the causes of plating defects. In a bid to tackle thisproblem, we will apply a casual reasoning model equipped with deep learning to find the underlyingcauses of process failures.

Développement et analyse d’algorithmes efficaces pour le Building Information Modeling 4D et applications à la recherche de chemins

Depuis quelques années, la nécessité d'avoir des plans numériques afin de gérer efficacement les espaces est devenue un enjeu de taille. La modélisation des données du bâtiment (Building Information Modeling ou BIM) permet non seulement d’obtenir des plans 3D de grande précision, mais aussi des données supplémentaires liées au bâtiment telles que l’évolution de l’espace dans le temps ou la géolocalisation des éléments.