Paris AI, ML, and Computer Vision Meetup - July 28, 2026

Join our in-person meetup to hear talks from experts on cutting-edge topics across AI, ML, and computer vision.

This talk will explore practical workflows for building, evaluating, and improving modern computer vision systems. We’ll dive into real-world approaches to dataset curation, model analysis, multimodal AI workflows, and production-ready vision pipelines using open-source technologies.

The session is designed for engineers, researchers, and AI practitioners looking to better understand how teams are developing and scaling computer vision applications today. Expect practical demos, technical insights, and discussions around the evolving AI tooling ecosystem.

Vision Transformers (ViT) dominate computer vision. However, their reliance on rigid patch projectors hinders transfer to Earth Observation (EO), where inputs vary widely in modality, scale, and resolution. We introduce UniverSat, a ViT-style backbone built around a Universal Patch Encoder that maps patches from arbitrary spatial, spectral, and temporal resolutions, and from both optical and non-optical sensors, into a shared embedding space with a single set of weights. This enables training one model on heterogeneous multimodal corpora in self-supervision, yielding robust sensor-agnostic spatial features.

State-of-the-art image generation models require massive web scrapes and expensive post-training alignment. This talk explores three recent works that challenge the "bigger is better" paradigm to build efficient and controllable models. First, we show how ImageNet alone (only 1/1000th the training data of Stable Diffusion) can match billion-scale models using a fraction of the compute. Second, we introduce a frequency-balanced training objective that overcomes spectral bias, learning high-fidelity textures up to 40% faster. Finally, we present MIRO, a multi-reward pretraining method that bakes human preferences directly into the model, bypassing costly post-hoc RLHF and outperforming models 30x its size.

Vision-language contrastive pretraining has become the dominant recipe for 3D radiology foundation models. However, medical images and reports are much more complex than their natural counterparts, and CLIP-style pretraining compresses this structure by encoding each modality into a single global token, at the risk of losing important details. We present ConQuer (Concept Queries), an image-text pretraining method that augments CLIP's global alignment with a set of localized alignments, one per concept. We use this method to train Jolia, a 3D CT foundation model for chest and abdominal CT. Jolia consistently outperforms a CLIP baseline on findings classification, report generation, and cross-center transfer, and sets a new state of the art across multiple public benchmarks.

From a single GPU server with SSH access to a Kubernetes platform supporting 250+ ML engineers — Rémi Debette, MLOps Manager at InstaDeep, shares the journey behind AIchor, InstaDeep's internal AI workloads platform. In this talk, Rémi covers the key architectural choices that made AIchor scalable, secure, and easy to use for researchers: GitOps-based experiment triggers, multi-cloud Control/Data Plane architecture, and automated infrastructure provisioning. He'll also share how the platform is now expanding beyond training to tackle inference at scale, bringing the same level of automation and observability to model serving.