This seminar aims to increase the links between the different laboratories in Saclay in the field of Applied Maths, Statistics and Machine Learning. The Seminar is organized every first Tuesday of the month with 2 presentations followed by a small refreshment. The localization of the seminar will change to accommodate the different labs.
Organization
Due to access restriction, you need to register for the seminar. A link is provided in the description and should also be sent with the seminar announcement. It will also help us organize for the food quantities. If you think you will come, please register! (even if you are unsure)
To not miss the next seminar, please subscribe to the announcement mailing list subscribe/palaisien.
You will receive email on the palaisien@inria.fr mailing list.
You can also add the calendar from the seminar to your own calendar (see below).
Next seminars
REGISTER
06 Oct 2026, 12h At Inria Saclay - Amphi Sophie Germain
Julie Alberge-SurvICL: In-Context Learning for Survival Analysis
Survival analysis aims to predict the time until an event of interest, such as death, disease progression, or equipment failure. Unlike standard regression, it must account for censoring, which occurs when the event has not been observed for some individuals, for example because they leave the study or the study ends. This makes survival prediction particularly challenging, especially on small datasets commonly encountered in clinical research. Recent advances in tabular foundation models...
Survival analysis aims to predict the time until an event of interest, such as death, disease progression, or equipment failure. Unlike standard regression, it must account for censoring, which occurs when the event has not been observed for some individuals, for example because they leave the study or the study ends. This makes survival prediction particularly challenging, especially on small datasets commonly encountered in clinical research.
Recent advances in tabular foundation models have demonstrated remarkable performance in classification and regression, often outperforming traditional methods such as gradient-boosted trees. However, their success has not yet translated to survival analysis, where the Cox proportional hazards model, introduced over 50 years ago, remains a strong baseline.
In this talk, we introduce SurvICL, a tabular foundation model that predicts individual survival curves through in-context learning, without dataset-specific training or hyperparameter tuning. SurvICL is pretrained entirely on synthetic datasets generated from a flexible proportional-odds prior and learns to predict survival curves using a simple quantile regression loss. We also present our benchmark of 63 survival datasets, the largest to date, and show how SurvICL outperforms 20 competing methods, including classical statistical models, tree-based approaches, deep learning methods, and other foundation models. In particular, SurvICL is the first tabular foundation model to outperform Cox on small survival datasets.
Enzo Tartaglione-Bringing Training to the Edge: Subspace and Sparsity for Efficient Deep Learning
Deep models keep growing, which makes them hard to deploy and even harder to train on edge devices. Classic pruning often fails to deliver real speed-ups. Unstructured sparsity is poorly supported by general-purpose hardware, and on parallel hardware, thinner layers do not shorten the critical path. In this talk, I first present ways to reduce network depth by collapsing layers, removing non-linearities so adjacent layers can be merged (EASIER, TLC, LaCoOT), and to reduce redundant tokens in ...
Deep models keep growing, which makes them hard to deploy and even harder to train on edge devices. Classic pruning often fails to deliver real speed-ups. Unstructured sparsity is poorly supported by general-purpose hardware, and on parallel hardware, thinner layers do not shorten the critical path. In this talk, I first present ways to reduce network depth by collapsing layers, removing non-linearities so adjacent layers can be merged (EASIER, TLC, LaCoOT), and to reduce redundant tokens in multimodal models (FOLDER). I then discuss why pruning at initialization remains hard, and how training itself can be made cheaper by freezing neurons that have reached equilibrium (NEq). Finally, I address memory, the main bottleneck for on-device learning. Because the optimization subspace is stable during fine-tuning, activations and weights can be compressed together through low-rank decomposition. This cuts backpropagation memory and FLOPs by orders of magnitude. I close with open questions on foundation models, continual learning, pruning versus quantization, and links and opportunities to other (more or less) related fields.
REGISTER
10 Nov 2026, 12h At Inria Saclay - Amphi Sophie Germain
Scientific Committee
The program and the organization of this seminar is driven by a scientific committee composed of members of the different laboratories in Saclay. The members of the committee are currently:
