Valentin Thomas
I am a Senior Machine Learning Scientist at Layer6 where I
mainly work of foundation models for tabular data and time series. I
previously graduated from my PhD at Mila, where I worked on
Reinforcement Learning and Deep Learning under supervision of
Yoshua Bengio and Nicolas Le Roux.
Email  / 
CV  / 
Bio  / 
Google Scholar  / 
Twitter  / 
Github
|
|
Research
I'm interested in reinforcement learning, deep learning and optimization. I have worked on unsupervised RL, planning, generalization in deep learning and optimization aspects of reinforcement learning. Representative papers are highlighted. Stars * indicate first authorship.
|
|
TabDPT: Scaling Tabular Foundation Models
Valentin Thomas*, Junwei Ma*, Rasa Hosseinzadeh,
Hamidreza Kamkari, Alex Labach, Jesse C. Cresswell,
Keyvan Golestan,
Guangwei Yu, Maksims Volkovs, Anthony Caterini
We introduce TabDPT, a tabular foundation model capable of providing predictions for unseen tabular datasets with no further training or hyperparameter tuning, and demonstrate scaling in both model and dataset size.
|
|
Retrieval and Fine-tuning for In-Context Tabular Models
Valentin Thomas*, Junwei Ma*, Rasa Hosseinzadeh, Keyvan Golestan,
Guangwei Yu, Maksims Volkovs, Anthony Caterini
NeurIPS 2024
ICML 2024 Workshop on In-Context Learning
We use retrieval with fine-tuning to improve in-context learning on tabular data and show large improvement with better scaling.
|
|
In-Context Data Distillation with TabPFN
Junwei Ma, Valentin Thomas, Guangwei Yu, Anthony Caterini
ICLR 2024 Workshop on Mathematical and Empirical Understanding of Foundation Models
(ME-FoMo)
We propose a simple method similar to prompt tuning to allow
in-context foundation tabular models to scale efficient with the
dataset size.
|
|
Bridging the Gap Between Target Networks and Functional Regularization
Alexandre Piché*, Valentin Thomas*, Joseph Marino, Rafael Pardinas, Gian Maria Marconi, Christopher Pal, Mohammad Emtiyaz Khan
TMLR 2023
NeurIPS 2021 DeepRL workshop
We analyze the implicit regularization performed by using Target Networks and show that, surprisingly, it can unstabilize TD. We propose a theoretically grounded alternative method, Functional Regularization, which alleviates these theoretical issues and performs well empirically.
|
|
On the role of overparameterization in off-policy Temporal Difference learning with linear function approximation
Valentin Thomas*
NeurIPS 2022
We study the role of overparameterization in Temporal Difference (TD) learning and how it affects optimization. For this, we analyze the spectrum of the Temporal Difference operator when using random features and with some assumptions of the Markov transition kernel.
|
|
The Role of Baselines in Policy Optimization
Jincheng Mei*, Wesley Chung, Valentin Thomas, Bo Dai, Csaba Szepesvari, Dale Schuurmans
NeurIPS 2022
Using value function baselines in on-policy stochastic natural policy gradients help achieve convergence toward globally optimal policy by reducing update aggressiveness rather than variance.
|
|
Beyond variance reduction: Understanding the true impact of baselines on policy optimization
Valentin Thomas*, Wesley Chung*, Marlos C. Machado, Nicolas Le Roux
ICML 2021
blog post/ICML talk
We show empirically and theoretically that despite common wisdom, baselines in policy gradient optimization have an effect beyond variance reduction and can impact convergence.
|
|
On the interplay between noise and curvature and its effect on optimization and generalization
Valentin Thomas*, Fabian Pedregosa, Bart van Merriënboer, Pierre-Antoine Mangazol, Yoshua Bengio, Nicolas Le Roux
AISTATS 2020
Oral talk at the 2020 Workshop on theory of deep learning at
the Institute for Advanced Studies, Princeton
AISTATS talk
We show how the interplay between the local curvature of the loss (the hessian) and the local gradient noise (the uncentered gradient covariance) can impact optimization and generalization in neural networks.
|
|
Planning with Latent Simulated Trajectories
Alexandre Piché, Valentin Thomas, Cyril
Ibrahim, Yoshua Bengio, Julien Cornebise and Chris Pal
ICLR 2019 Workshop on Structure and Priors in Reinforcement Learning
Extension of our work "Probabilistic Planning with Sequential Monte Carlo methods" by treating the trajectory as a latent variable and using an EM algorithm.
|
|
Probabilistic Planning with Sequential Monte Carlo methods
Valentin Thomas*, Alexandre Piché*, Cyril
Ibrahim, Yoshua Bengio and Chris Pal
ICLR 2019
Contributed talk at NeurIPS 2018 workshop Infer to Control
Leveraging control as inference and Sequential Monte Carlo methods, we propose a probabilistic planning algorithm.
|
|
Disentangling the independently controllable factors of variation by interacting with the world
Valentin Thomas*, Emmanuel Bengio*, William Fedus*, Jules Pondard, Philippe Beaudoin, Hugo Larochelle, Joelle Pineau,
Doina Precup and Yoshua Bengio
Oral at NeurIPS 2017 workshop on Learning Disentangled Representations: from
Perception to Control
We draw a connection between mutual information and the intrinsic reward function (through the Donsker-Varadhan representation of the Kullback-Leibler divergence) used for learning jointly options/factors and latent representations in Independently Controllable Factors.
|
|
Independently Controllable Factors
Valentin Thomas*, Jules Pondard*, Emmanuel Bengio*, Marc Sarfati, Philippe Beaudoin, Marie-Jean Meurs, Joelle Pineau, Doina Precup, Yoshua Bengio
Presented at the Montreal AI Symposium 2017
This work is a finalized version of Independently Controllable Features where the policies and factors are now embedded in a contiuous space. We demonstrate how one can use the features learnt.
|
|
Independently Controllable Features
Emmanuel Bengio*, Valentin Thomas, Joelle Pineau, Doina Precup, Yoshua Bengio
RLDM 2017
We propose a way to learn jointly a set of discrete policies each affecting a component of the latent state representation for unsupervised reinforcement learning. We hypothesize that this process discovers controllable factors of variation in the world as well as how to control them.
|
|
Decoupling Backpropagation using Constrained Optimization Methods
Valentin Thomas*, Akhilesh Gotmare*,
Johanni Brea and Martin Jaggi
In ICML 2018 workshop on Efficent Credit Assignement
We propose BlockProp which lets one train deep neural networks in model parallel fashion, where parts of the model may reside on different devices (GPUs).
|
|
Preserving the entanglement of two qubits with feedback control
Valentin Thomas*, Pierre Rouchon
Report for a research semester in 2014 (in french)
This research project was about designing a feeback control loop using an electromagnetic field to preserve the entanglement of two qbits. This is necessary as because of quantum decoherence the entanglement tends to vanish which is a major issue in developping quantum computer hardware. We proposed a simple Lyapunov-based feedback control loop.
|
Experience
Here you can find the internships I have done during my MSc at Mines Paris and at Ecole Normale Superiéure Paris-Saclay then during my PhD at University of Montreal.
|