About the seminar

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 palaisien@inria.fr.
You can also add the calendar from the seminar to your own calendar (see below).

Next seminars

REGISTER 11 Feb 2020, 16h At Telecom Paris - Amphi 2
Hicham Janati - Spatio-temporal Optimal transport metric for brain imaging data
Evaluating how two different MEG time series are close to each other is far from trivial. In this paper, we propose Spatio-Temporal Alignments (STA), a new differentiable formulation of DTW, in which spatial differences between time samples are accounted for using regularized optimal transport.
Comparing data defined over space and time is notoriously hard, because it involves quantifying both spatial and temporal variability, while at the same time taking into account the chronological structure of data. For instance, MEG and EEG source estimates provide brain activity maps in a millimetre / millisecond resolution with significant spatio-temporal differences. Evaluating how two different MEG time series are close to each other is thus far from trivial. This work is a first step towards that goal. Dynamic Time Warping (DTW) computes an optimal alignment between time series in agreement with the chronological order, but is inherently blind to spatial shifts. In this paper, we propose Spatio-Temporal Alignments (STA), a new differentiable formulation of DTW, in which spatial differences between time samples are accounted for using regularized optimal transport (OT). Our temporal alignments are handled through a smooth variant of DTW called soft-DTW, for which we prove a new property: soft-DTW increases quadratically with time shifts. The cost matrix within soft-DTW that we use is computed using unbalanced Optimal transport. Experiments on brain imaging data confirm our theoretical findings and illustrate the effectiveness of STA as a dissimilarity for spatio-temporal data.
Rainer Dyckerhoff - Convergence of depths and central regions
After a short introduction in the general concept of depth we first discuss the continuity properties of depths and central regions. The main part of the talk is concerned with the connections between different types of convergence. We give conditions under which the pointwise (resp. uniform) convergence of the data depth implies the pointwise (resp. uniform) convergence of the central regions in the Hausdorff metric as well as conditions under which the reverse implications hold.
Depth is a concept that measures the centrality of a point in a given data cloud x_1, x_2,..., x_n in ℝ^d or in a given probability distribution on ℝ^d. Every depth defines a family of so-called central regions. The α-central region is the set of all points that have a depth of at least α. For statistical applications it is desirable that with increasing sample size the empirical depth as well as the empirical central regions converge almost surely to their population counterparts.

After a short introduction in the general concept of depth we first discuss the continuity properties of depths and central regions. The main part of the talk is concerned with the connections between different types of convergence. We give conditions under which the pointwise (resp. uniform) convergence of the data depth implies the pointwise (resp. uniform) convergence of the central regions in the Hausdorff metric as well as conditions under which the reverse implications hold. Further, we demonstrate that under relative weak conditions the pointwise convergence of the data depth (resp. central regions) is equivalent to the uniform convergence of the data depth (resp. central regions). Finally, we illustrate these results by applying them to special notions of data depth that have been proposed in the literature.
REGISTER 03 Mar 2020, 16h At INRIA Saclay - Batiment Alan Turing - Amphi sophie Germain
Alexei Grinbaum - Chance as a value for artificial intelligence
Deep learning techniques lead to fundamentally non-interpretable decisions made by the machine. Although such choices do not have an explanation, they impact the users in significant ways. If the ultimate innovator is a machine, what is the meaning of responsible conduct?
Deep learning techniques lead to fundamentally non-interpretable decisions made by the machine. Although such choices do not have an explanation, they impact the users in significant ways. If the ultimate innovator is a machine, what is the meaning of responsible conduct? I argue in a recent book that the capacity to extract an AI system from human judgment, by reducing transparency in favor of opacity, is an essential value in machine ethics. This can be achieved through the use of randomness, illustrated on several examples including the trolley dilemma. Methodologically, a comparison of common motives between technological setups and mythological narratives is used to achieve ethical insights.
Yannig Goude - Machine Learning Methods for Electricity Load Forecasting: contributions and perspectives.
To maintain the electricity quality, energy stakeholders are developing smart grids, the next generation power grid including advance communication networks and associated optimisation and forecasting tools. We will present recent development in the field of online learning and probabilistic forecasting done at EDF in this context.
Energy systems are facing a revolution and many challenges. On the one hand, electricity production is moving to more intermittency and complexity with the increase of renewable energy and the development of small distributed production units such as photovoltaic panels or wind farms. On the other hand, consumption is also changing with e.g. plug-in (hybrid) electric vehicles, heat pumps, the development of new technologies such as smart phones, computers, storage devices. To maintain the electricity quality, energy stakeholders are developing smart grids, the next generation power grid including advance communication networks and associated optimisation and forecasting tools. Exploiting the smart grid efficiently requires advanced data analytics to improve forecasting at different geographical scale. We will present recent development in the field of online learning and probabilistic forecasting done at EDF in this context.

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:

Funding

This seminar is made possible with financial support of the ENSAE and DataIA.