Thibaut Durand

I'm a researcher driven to apply Machine Learning (ML) research to real-world problems. I currently work as a Staff Machine Learning Researcher at RBC Borealis (formerly Borealis AI) in Vancouver, BC, Canada, where I work at the intersection of ML and finance. Previously, I was a Postdoctoral Fellow in Computer Vision and Machine Learning at Simon Fraser University, working under the supervision of Greg Mori. I earned my PhD from Sorbonne University in Paris, France, where I was advised by Matthieu Cord and Nicolas Thome. My thesis was focused on weakly supervised learning for visual recognition and was recognized with PhD awards from AFRIF and DGA. I regularly serves as an Area Chair or reviewer at major computer vision and machine learning conferences, including CVPR, ICCV, ECCV, ICLR, NeurIPS, and ICML. I'm French and Swiss, and my name is pronounced Tibo.

Email  /  CV  /  Bio  /  Scholar  /  GitHub  /  LinkedIn

Experience

I'm a Staff Machine Learning Researcher at RBC Borealis, which is the AI research arm of the Royal Bank of Canada (RBC). RBC is ranked 3rd globally and 1st in Canada in the Evident AI Index (published in October 2024), which evaluates the AI adoption and maturity throughout the largest banks in North America, Europe, and Asia. I work on revolutionize finance by creating intelligent and human-centered systems that drive meaningful change to the future of banking. Since joining RBC Borealis, I have worn multiple hats, serving as an independent contributor, technical lead, and project lead. Below is a summary of the main leadership and independent contributor skills I have developed and continue to strengthen.

Leadership skills

• I have experience leading end-to-end ML projects, from scoping and translating complex business requirements into technical solutions, to aligning cross-functional stakeholders, building MVPs, and deploying production-ready models.
• I led cross-functional project teams of 5-8 researchers and engineers to design, develop, and deploy end-to-end AI solutions, driving collaboration across domains and ensuring alignment with business objectives, technical requirements, and delivery timelines.
• I have experience managing competing priorities by planning and executing research sprints that balance short-term deliverables with long-term innovation goals. I focus not only on meeting project milestones but also on maintaining team health and codebase quality, recognizing that sustainable technical progress is closely tied to team well-being and a healthy engineering foundation.
• I developed long-term ML research roadmap. I led the development of the Asynchronous Temporal Models (ATOM) and Responsible AI research roadmaps, shaping foundation model strategy for financial services and guiding priorities for trustworthy, human-centric AI.
• I mentored junior ML researchers to support their growth in both technical expertise and leadership capabilities, providing guidance on research direction and cross-functional collaboration. I also mentored software engineers on ML concepts and best practices, helping them build a strong foundation in applied ML and contribute more effectively to AI-driven projects.

Independent contributor skills

• I have expertise in designing and implementing ML solutions to address real-world challenges, with a strong focus on aligning technical implementation with strategic business goals and practical constraints. I have developed innovative approaches that push the boundaries of conventional solutions while ensuring real-world applicability and impact.
• I have experience developing ML models across a wide range of applications, including visual recognition, recommendation systems, user representation, personalization, asynchronous time series modeling, time series forecasting, and risk assessment. My work also includes the development of explainable deep learning models to enhance transparency and trust.
• I have worked on developing evaluation frameworks to assess ML model performance across critical business segments, ensuring alignment between ML metrics and measurable business outcomes. I also designed metrics for both offline evaluations and online experiments to support data-driven decisions.
• I have experience building ML datasets for real-world data, encompassing feature engineering, feature selection, data quality, and data preprocessing techniques such as feature scaling and handling missing values. My experience primarily involves working with image, time series, and tabular datasets, with additional exposure to other modalities such as text and molecular data.
• I primarily develop ML models using PyTorch, leveraging its flexibility and efficiency for both research and production workflows. I have experience harnessing GPU accelerators and advanced distributed training techniques to efficiently scale training for large and complex workloads.

Collaboration & Teamwork

• I have experience working with complex, distributed systems where individual components are owned and maintained by different teams. I collaborated with diverse teams to align on architecture, resolve integration challenges, and successfully deploy ML models into existing legacy infrastructures.
• I have experience fostering connections and building trust among teams and individual members to strengthen collaboration and drive shared success.
• I have experience working in geographically distributed teams across multiple office locations, collaborating effectively across time zones to ensure smooth communication and alignment.
• I quickly adapt to new challenges and actively seek opportunities to grow by learning new concepts, developing new skills, and exploring innovative ideas. Throughout my career, I’ve worked on a wide range of tasks and data types, collaborating effectively with diverse teams across various domains.
• I have experience collaborating on and contributing to shared codebases, with a focus on building maintainable code and managing technical debt to ensure long-term efficiency and scalability.

Research

My research focuses on machine learning and deep learning, with a particular interest in weakly supervised learning, temporal modeling, and multimodal learning. These areas are key to building human-centred AI systems that understand the complex and dynamic nature of the real world. I have a background in computer vision, but I enjoy exploring new modalities and currently work extensively with asynchronous time series, also known as event-based data.

Below are selected publications presented at top-tier conferences and journals. For the full list, please visit my Google Scholar profile.

LAST SToP is an efficient method to adapt LLMs for asynchronous time series while preserving semantic information through language.

Variational Selective Autoencoder is a model designed to learn from partially observed heterogeneous data by selectively encoding observed features and effectively handling missing information.

A model that learns user-conditioned representations to improve open vocabulary image hashtag prediction by capturing individual user preferences and context.

LayoutVAE is a generative model that learns to produce diverse and coherent scene layouts from a given set of object labels using a variational auto-encoder framework.

A variational auto-encoder framework for modeling stochastic point processes by learning flexible representations of temporal event data using neural networks.

A deep ConvNet approach for multi-label classification that can learn effectively from partially labeled data by leveraging label correlations and handling label uncertainty during training.

A deep latent structured model that incorporates negative evidence through a novel pooling strategy, enabling more accurate classification, ranking, and weakly supervised segmentation by explicitly penalizing incorrect class predictions.

WILDCAT is a weakly supervised ConvNet that, using only global image labels, learns multiple class-specific feature maps plus a novel min/max pooling mechanism to enable image classification, point-wise localization, and semantic segmentation.

WELDON is a weakly supervised deep learning approach that improves image classification and localization by selecting and aggregating both the most and least activated regions in convolutional feature maps using a novel pooling strategy.

MANTRA is a latent structural SVM framework that optimizes both minimum and maximum scoring regions in images to improve weakly supervised image classification and ranking tasks.

Design and source code from Jon Barron's website