As Artificial Intelligence (AI) is gaining widespread use, Layer 6 is using the technology to solve real problems and directly shape how teams work, models are built, and innovation is delivered at scale across the Bank. Layer 6 and TD partner with the independent, non-profit AI research centre Vector Institute as a Platinum and Founding Sponsor to make progress in areas that matter most to academia and the financial services industry, with direct access to Vector faculty and leadership.
Turning Research into Scalable AI Outcomes
Central to our partnership are collaborative research efforts between scientists at Layer 6, and Vector faculty research labs. These collaborations are born out of the challenges of deploying AI in real-world applications across the Bank, including the need to ensure AI is used responsibly.
Read about how Layer 6 improves trustworthiness of AI
One such example is the recent research collaboration with Vector faculty member Rahul Krishnan to build the first causal foundation model. Causal modelling is a branch of AI that tries to predict what “would have happened” if a scenario was slightly altered. This is extremely tricky to get right, as it is not possible to perfectly replay a scenario. Still, when a causal model is effective, it allows the user to reason about alternative decisions and make the best choices.
Typical approaches to causal modelling involve domain experts studying the data by hand, and using their judgement to propose relationships between variables. The downside is that for every new problem, domain experts have to start over from scratch. When deploying AI at scale, development cycles need to be rapid and models need to be extensible, so the manual and iterative nature of causal modelling limits its impact.
Layer 6 and Dr. Krishnan’s group set out to develop a new paradigm for causal modelling, inspired by advances in foundation models. In other branches of machine learning, like natural language processing, foundation models power some of the most advanced AI that is now seeing widespread use in chat applications. Foundation models are large-scale neural networks trained on vast quantities of data to perform general-purpose tasks, acting as a core knowledge resource. Then, when a new task is identified, foundation models can be quickly adapted to leverage their general knowledge for specific outcomes.
See how Layer 6 builds foundation models for tabular data
The research collaboration designed the first causal foundation model – a large scale model trained on many hypothetical “what-if” scenarios to learn the types of relationships that can arise between variables. Then, when a new causal question comes up, the foundation model is able to use examples it has seen before and its collected knowledge to replay altered scenarios without any further training. This flips the script on how new causal models are developed – rather than a domain expert studying the data by hand, any new dataset can be fed to the foundation model to predict causal effects instantly. Importantly, many different applications can be handled by a single foundation model, enabling the technology to scale to applications across the enterprise with much less overhead.
Read the full research work on causal foundation models
In the collaboration, Vector researchers brought deep theoretical knowledge about causal modelling, and Layer 6 scientists repurposed their experience in training large-scale foundation models. These complementary skillsets were crucial for the overall vision to come together.
Empowering Teams Through Hands-On Learning
The Vector Institute ecosystem involves much more than just cutting-edge research. They regularly host training opportunities that Layer 6 and TD scientists participate in both as presenters and participants, including webinars, bootcamps, and applied AI workshops.
In a recent Interpretability and Explainability Bootcamp, a cross-functional team from TD spanning engineers, scientists, and product owners worked together to investigate explainability in applications of large language models (LLMs). These models, which involve many billions or trillions of artificial neurons, are too complex for humans to comprehend entirely (after all, the human brain has a mere 86 billion neurons). But, when used for important applications at the Bank, LLMs still need to be explainable – we must be able to provide reasoning for the outputs and predictions that they make. The bootcamp team looked at chatbot applications of LLMs and ways to enhance their transparency. The hands-on sessions, led by Vector faculty and researchers, equipped teams with new approaches they could immediately test on real-world AI systems.
Recognizing and Building Talent
Through its deep integration with top universities all across Canada, the Vector Institute is a talent magnet and incubator. Layer 6 has partnered with many Vector Institute faculty members through the University of Toronto’s MScAC program to provide research-based internships to local students. These projects provide students with hands-on experience working in a fast-paced research environment, where projects are driven by curiosity and practical needs. Today’s students power tomorrow’s AI innovations and through this partnership Layer 6, Vector, and UofT contribute to not only high-quality research outcomes, but the training of future AI leaders.
View all of Layer 6’s past MScAC projects
A Research Partnership with Lasting Value
Together, Layer 6 and Vector Institute researchers have co-authored a growing body of research that has gained attention in the wider AI community. These published outcomes not only contribute to academic advancement of knowledge, but also serve as a foundation for Layer 6’s evolving AI strategy.
Key publications from years of collaboration include:
- A Geometric Explanation of the Likelihood OOD Detection Paradox
Solves a longstanding paradox around how generative models recognize data different from what they saw during training. - MultiResFormer: Adaptive Time Series Forecasting
Adapts the transformer architecture for time series problems by identifying periodic behaviours at multiple resolutions. - Augment then Smooth: Reconciling Differential Privacy with Certified Robustness
Combines two different aspects of Trustworthy AI, namely privacy and robustness, to give future-proof guarantees on reliability. - TR0N: Translator Networks for 0-Shot Plug-and-Play Conditional Generation
Enables unconditional generative models to preform conditional generation with no need for additional training data or fine-tuning. - DuETT: Dual Event Time Transformer for Electronic Health Records
Combines time-series and event information from health records into a powerful and efficient architecture for health-focused AI. - X-Pool: Cross-Modal Language-Video Attention for Text-Video Retrieval
Introduces a cross-modal attention mechanism that enables text queries to attend directly to relevant video frames. - Bayesian Nonparametrics for Offline Skill Discovery
Presents a method for reinforcement learning agents to discover a dynamic set of skills through variational inference. - C-Learning: Horizon-Aware Cumulative Accessibility Estimation
Advances reinforcement learning with value estimation that is horizon-aware for trading off speed and reliability.
Contributing to the Broader AI Community
The partnership has also created space for Layer 6 colleagues to give back to the broader ecosystem. Layer 6 researchers regularly mentor students in the Vector network, participate in career fairs, and share insights through talks and panels hosted by Vector. Through these channels, TD helps bridge the gap between industry and academia—fueling a pipeline of talent and ideas that benefit Canada’s AI landscape as a whole.
Listen to a Layer 6 seminar hosted by Vector Institute
Through our research and ecosystem partnerships, Layer 6 is committed to continue shaping how AI is used by ensuring that the products and systems we build are safe and designed to protect our customers.
