Paris Noah's Ark Lab

We are a research laboratory located in Paris, working on the fundamental and practical development of modern artificial intelligence systems. Paris Noah's Ark Lab consists of 3 research teams that cover the following topics:

• Time Series and Transfer Learning (Team leader: Ievgen Redko)
• Auto-Data-Science and Reinforcement Learning (Team leader: Balazs Kegl)
• Autonomous Driving

Projects

Packages

• Mantis: Lightweight Calibrated Foundation Model for User-Friendly Time Series Classification: check also our paper.

Preprints

• Time Series Representations for Classification Lie Hidden in Pretrained Vision Transformers: converting time series to images and feeding them into a pre-trained ViT.
• TAG: A Decentralized Framework for Multi-Agent Hierarchical Reinforcement Learning: distributed multi-agent hierarchical reinforcement learning framework.
• SKADA-Bench: Benchmarking Unsupervised Domain Adaptation Methods with Realistic Validation: benchmark of shallow and deep domain adaptation method with realistic validation
• Clustering Head: A Visual Case Study of the Training Dynamics in Transformers: visual and theoretical understanding of training dynamics in transformers.
• Large Language Models as Markov Chains: theoretical insights on their generalization and convergence properties.
• A Systematic Study Comparing Hyperparameter Optimization Engines on Tabular Data: insights to navigate the maze of hyperopt techniques.

2025

• (ICML'25 FMSD Workshop, Best Paper) CauKer: Classification Time Series Foundation Models Can Be Pretrained on Synthetic Data Only: algorithm to generate diverse synthetic time series.
• (ICML'25) AdaPTS: Adapting Univariate Foundation Models to Probabilistic Multivariate Time Series Forecasting: simple yet powerful tricks to extend foundation models.
• (ICLR'25) Zero-shot Model-based Reinforcement Learning using Large Language Models: disentangled in-context learning for multivariate time series forecasting and model-based RL.
• (ICASSP'25) Easing Optimization Paths: A Circuit Perspective: mechanistic study of training dynamics in transformers.
• (Neurocomputing) Self-training: A survey: know more about pseudo-labeling strategies.

2024

• (NeurIPS'24) MANO: Unsupervised Accuracy Estimation Under Distribution Shifts: when logits are enough to estimate generalization of a pre-trained model.
• (NeurIPS'24, Spotlight) Analysing Multi-Task Regression via Random Matrix Theory: insights on a classical approach and its potentiality for time series forecasting.
• (ICML'24, Oral) SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting: sharpness-aware minimization and channel-wise attention is all you need.
• (AISTATS'24) Leveraging Ensemble Diversity for Robust Self-Training: confidence estimation method for efficient pseudo-labeling under sample selection bias.
• (JMLR, 2024) Multi-class Probabilistic Bounds for Majority Vote Classifiers with Partially Labeled Data: generalization with unlabeled or pseudo-labeled data.
• (ICML '24) Position: A Call for Embodied AI: position paper on the need for embodied AI research
• (RLC '24) A Study of the Weighted Multi-step Loss Impact on the Predictive Error and the Return in MBRL: multi-step loss in MBRL does not work as well as expected

2023

• (ICML '23) Meta Optimal Transport
• (AAAI '23) Unbalanced Co-Optimal Transport
• (ICML '23) Multi-Agent Best Arm Identification with Private Communications
• (ICML '23) Random Matrix Analysis to Balance between Supervised and Unsupervised Learning under the Low Density Separation Assumption
• (ICML '23) PCA-based Multi Task Learning: a Random Matrix Approach