Joint Meetup Sicara x Mobiskill

👋 The Next Meetup will be on Wednesday, June 22th in a hybrid format with Mobiskill

As usual, there will be two talks on the program. Each talk will be followed by a question and answer session with the speaker.

🗓 The program:

Post mortem of a failed computer vision launch by Mateo Rojas Carulla

From the streets and hospitals to our homes, machine learning systems are already pervading all aspects of our lives. At the same time, it fails a lot, challenging how much we can trust these new technologies. When building production-level systems, computing aggregate metrics on a test dataset is no longer sufficient to ensure quality. Why is it so challenging to thoroughly test machine learning systems before they’re deployed in the real world? How can we generally bring machine learning from an academic endeavour to a rigorous engineering discipline?

Automatically remove outliers in support sets by Tarek Ayed

Few-shot learning workflows rely on a very small number of examples of each class. This means that the quality of the support set (the set of labelled examples) is crucial for the model's performance. In fact, a FSL model's performance can drop by 1/3 with 20% of the support set being outliers.
I will present a simple and fast baseline for outlier detection in support sets which allows to catch 80% of outliers with only 35% of false positives.