Hugo Larochelle

Hugo Larochelle

Mila - Quebec Artificial Intelligence Institute · Université de Montréal · McGill University
7 previous affiliations
145Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Hugo Larochelle focuses on deep learning architectures, generative modeling, and few-shot learning methodologies. His early work established foundational techniques for training deep neural networks and autoregressive density estimators. More recently, his research addresses the evaluation and alignment of large language models, alongside applications in ecological monitoring and multi-agent forecasting.

Research Trajectory

The trajectory begins with foundational contributions to the training of deep neural networks, addressing the optimization challenges of highly varying functions. In Greedy Layer-Wise Training of Deep Networks, he explored unsupervised pre-training regimes to overcome the limitations of shallow kernel machines. This work laid the groundwork for subsequent investigations into Restricted Boltzmann Machines (RBMs) and Deep Belief Networks. He demonstrated that RBMs could serve as competitive standalone non-linear classifiers in Classification using discriminative restricted Boltzmann machines, and further formalized greedy layer-wise unsupervised learning procedures in Exploring Strategies for Training Deep Neural Networks. A major architectural advancement followed with Stacked Denoising Autoencoders: Learning Useful Representations in a Deep Network with a Local Denoising Criterion, which introduced a method for learning robust, high-level features by training networks to reconstruct clean data from partially corrupted inputs.

Expanding into structured and sequential data processing, his research investigated the area of Foveated Architectures for Efficient and Goal Directed Vision. In Learning to combine foveal glimpses with a third-order Boltzmann machine, he developed a model capable of accumulating information over a sequence of fixations, mimicking retinal glimpses to process high-resolution images efficiently. Concurrently, he made significant contributions to the research direction of Autoregressive Density Estimation with Neural Networks. He introduced the Neural Autoregressive Distribution Estimator in The Neural Autoregressive Distribution Estimator, providing a tractable approach to modeling complex distributions and advancing work in the area of Learning and Inference in Exponential Output Spaces. This line of inquiry into the broader area of Variable Ordering in Neural Autoregressive Models culminated in MADE: Masked Autoencoder for Distribution Estimation, which presented a simple yet effective modification to autoencoders for estimating probability distributions by masking internal connections.

Transitioning toward model generalization and evaluation, he participated in the development of adversarial alignment techniques for unsupervised domain adaptation in Domain-Adversarial Training of Neural Networks. Recognizing the critical need for rigorous evaluation standards, he also contributed to community-wide efforts to ensure reliable research workflows, as detailed in Improving Reproducibility in Machine Learning Research(A Report from the NeurIPS 2019 Reproducibility Program). A central focus during this period was few-shot classification, aiming to recognize unseen classes from minimal examples across diverse domains. He proposed dynamic feature composition techniques in A Universal Representation Transformer Layer for Few-Shot Image Classification. To address diverging evaluation norms in the field, he helped establish unified protocols and benchmarks, introducing the VTAB+MD protocol in A Unified Few-Shot Classification Benchmark to Compare Transfer and Meta Learning Approaches and formalizing a large-scale evaluation framework in Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples.

Most recently, his work has shifted toward the capabilities, alignment, and evaluation of large language models (LLMs). As context windows expanded, he investigated the performance gains of providing thousands of examples in Many-Shot In-Context Learning, demonstrating that models can override pre-training biases without weight updates. He has also explored multi-agent systems, replacing human expert panels with LLMs for structured forecasting in DeLLMphi: A Multi-Turn Method for Multi-Agent Forecasting. In the domain of model alignment, he developed methods to specialize reward models using linear combinations of general features in Capturing Individual Human Preferences with Reward Features, and participated in research on token-level adversarial detoxification in REPO: Detoxifying LLMs via Representation Erasure-based Preference Optimization.

Beyond core LLM research, he has applied machine learning to complex environmental and economic systems. He evaluated the Segment Anything Model for monitoring reforestation efforts in Assessing SAM for Tree Crown Instance Segmentation from Drone Imagery, contributing to broader research directions such as Multimodal Environmental Fusion for High Resolution Ecological Mapping and Joint Species Distribution Modeling for Biotic Interactions. He also collaborated on applying reinforcement learning to climate-finance simulations, using opponent shaping to encourage sustainable cooperation in Towards Sustainable Investment Policies Informed by Opponent Shaping. Finally, addressing the need for interpretable AI benchmarks, he introduced a psychometric framework in BRIDGE: Predicting Human Task Completion Time From Model Performance, mapping abstract model scores to human-interpretable task completion times.

Current Directions

Recent work investigates the evaluation, alignment, and practical deployment of large language models and autonomous agents. A primary focus is on understanding and steering LLM behavior, whether by leveraging extended context windows for many-shot in-context learning, capturing highly individualized human preferences for reward modeling, or erasing toxic representations to improve model safety. Additionally, current research explores the psychometric evaluation of AI systems, seeking to translate abstract benchmark performance into interpretable metrics like human task completion time. Finally, these methodological advances are being applied to complex, real-world domains, including multi-agent forecasting, sustainable climate-finance simulations, and high-resolution ecological monitoring.

Profile updated

See something wrong? Report a mistake