Aurélie is starting by discussing the possibilities that machine learning and data science are opening, and mitigates them with the enormous ethical and political challenges we face.

Aurélie argues we need more women (and women of color and diverse backgrounds), and illustrates her talk with portraits of women she interviewed for her podcast.

All these amazing women live in a world where 84% of VC funding goes to all-men projects.

Aurélie encourages us to start our own companies (you never know whether you cannot do it until you've done it), or to choose wisely where we work, and more importantly to support and mentor each other. #sorority

#ParisWiMLDS

Julie now introduces competing risks analysis, a special type of survival analysis for when there are several evens of interest and you want to predict not only time-to-event, but the most likely event among the competing ones.

With her co-authors, Julie proposed a new approach for the competing risk settings, that scales to large amounts of data.

The underlying idea is to design a proper censoring-adjusted separable scoring rule that can be plugged into a multi-class classifier.

In her setting, the data is tabular, so unsurprisingly the best classifier is gradient boosted trees.

The approach has the best runtime to performance trade-off.

The work appeared at AISTATS'25 https://proceedings.mlr.press/v258/alberge25a.html and is implemented in a Python library hazardous.

#ParisWiMLDS

Julie introduces Inverse Propensity-of-Censoring Weighting (IPCW), which consists in re-weighting non-censored individuals with a weight that is inversely proportional to the probability of being censored.

This leads her to introduce two classical metrics of evaluation of the quality of the prediction of survival: the Briar score (equivalent to the MSE) and the C-index (equivalent to the ROC-AUC score).

#ParisWiMLDS

Julie starts by explaining the issue of censoring in survival data: what do you do with samples that did not experience the event of interest during the course of your study.

#ParisWiMLDS

Another example is face tracking, using the OpenCV tools for object detection.

Irina shows that although face detection works well, you need some adjustments (parameters tuning, conversion to grayscale, etc) to make smile/eye checking work.

Irina shows how she can use that for
- fun smile-triggered face masks (such as adding sunglasses)
- face (or conversely, background) blurring
- selecting portraits where the person has their eyes open
- and more

#ParisWiMLDS

Irina now describes a side project she worked on (she strongly advocates for working on side projects, to strengthen your skills and find what you enjoy working on), implementing a panorama function, stitching images together by extracting and matching descriptors.

Here she used SIFT descriptors, but OpenCV also has more modern descriptors based on deep learning, including SuperGlue
This approach is not just used in your phone camera, but also in robotics, medical imaging or 3D reconstruction from several 2D views.

#ParisWiMLDS