Biography
With a Bachelor's degree in Computer Science, I began my professional journey as a software developer in 2004. After a decade of software engineering, I pursued my longstanding fascination with the nature of intelligence by transitioning into machine learning. This led me to complete a Master's degree in Computer Science with a specialization in Artificial Intelligence (with a 4.24/4.3 GPA).
My thesis research focused on sample-efficient spatial planning using learned transition models in the context of model-based reinforcement learning. Since 2022, I have been deeply engaged with ARC-AGI, which I consider to be one of the most important challenges to AGI. My primary research interests lie in reasoning and generalization; fundamental challenges that continue to push the boundaries of current machine learning capabilities.
Currently, I serve as Senior Machine Learning Specialist at Trabotyx, an innovative AgriTech startup developing autonomous precision weeding robots to revolutionize sustainable agriculture.
Career highlights and progression:
-
Software Developer Era:
- [2004-2009] Lead developer for non-destructive testing systems, specializing in signal processing and visualization for ultrasonic, laser triangulation, and eddy current sensors.
- [2010] Three-month contract: Software optimization specialist at Bombardier, streamlining military aircraft design processes.
- [2010-2012] Senior software developer at Bloomberg LP, New York, developing critical financial infrastructure systems.
- [2013] Nine-month contract: Technical architect for Quebec government, designing enterprise-scale transactional web infrastructure.
- [2013-2015] Backend architect at a SaaS startup, developing enterprise collaboration platform infrastructure.
- [2015-2016] Founder of Nabla Analytics, developing a Big Data symbolic regression platform for quantitative finance. This venture catalyzed my transition into machine learning.
- [2017-2023] Machine learning applied researcher at a proprietary trading firm in Montreal, developing algorithmic trading systems. Concurrently completed a Master's degree and served as AI advisor at Trabotyx.
- [2022-current] Research focus on solving the ARC-AGI challenge, exploring fundamental aspects of machine intelligence and generalization.
- [2023-current] Transitioned from part-time AI Advistor to full-time Senior Machine Learning Specialist at Trabotyx, developing autonomous weeding robots to make organic farming more accessible.
Machine Learning Era:
Publications & Essays
2024
Towards Efficient Neurally-Guided Program Induction for ARC-AGI Runner-Up Paper Award at ARC Prize 2024
An analysis of the generalization capabilities of different approaches to neurally-guided program induction, applied to ARC-AGI.
No AGI Without Inference-time Search Montreal AI Symposium 2024
A position paper arguing that a search component at inference time is necessary for AGI, with learning-only algorithms being insufficient. View poster
Conviction-Based Planning for Sparse Reward Reinforcement Learning Problems ICAPS 2024 PRL Workshop
An analysis of the sample efficiency of various approaches to learning to solve planning problems.
2023
The Hitchhiker's Guide to the ARC Challenge 1st Place - Lab42 Essay Competition
An essay on algorithmic generalization in neural networks. Full PDF | Executive Summary
Counting and Algorithmic Generalization with Transformers arXiv Pre-print
This paper supports the idea that Transformers cannot learn to count in any true sense, including preliminary experimentation on potential generalization issues associated with layer normalization.
2022
Building human-like intelligence: an evolutionary perspective 1st Place - Lab42 Essay Competition
A conceptual essay on the principles of intelligence and the importance of the evolutionary perspective. Download PDF
Open Source Projects
Companion code to my paper "Towards Efficient Neurally-Guided Program Induction for ARC-AGI"
A data generation framework to experiment on program induction for ARC-AGI, and some visualization/grid manipulation utilities.
These scripts are the companion code to my paper "Counting and Algorithmic Generalization with Transformers".
A capital markets microstructure data simulation/generation framework (Python)
Source code related to a talk I've given about Bayesian generative modelling of stock returns. Implements some of the classical stochastic volatility models.
Contact
Email: s [dot] ouellette1 [at] gmail [dot] com
Social Media: