#43 Paris Women in Machine Learning & Data Science: Ethics, image and video surveillance
On June 14, we hosted our 43rd meetup. This was the first event since the addition of two new team members: Jihane and Juliette. We are so happy they joined the team!
Our meetup was held at Botify’s downtown Paris office, and they were an excellent host!
Our co-founder Caroline Chavier moderated the event and introduced the main theme of the meeting: ethics, image and video surveillance.
This topic is particularly relevant, as France recently announced that video surveillance will be used at the 2024 Olympics. This development raises valid concerns about potential consequences as it is possible that video surveillance will be implemented after the Olympics as well.
Caroline presented mission and history of WiMLDS Paris, and highlighted a few projects:
- First, she introduced the Research paper club, an online community to read, share and discuss ML and AI papers.
- Then, she introduced the Women in Data / AI festival, which will take place in Berlin, on June 30. Promotional tickets are available for the WiMLDS community.
- She also presented an event organised by one of our former hosts, Datacraft: The AI night (La nuit de l’IA). On June 26, they will award the best projects in data and AI and they are particularly focused on diversity and inclusion.
Our first speaker was Julie Delon, an applied mathematician, Lecturer and Researcher at Université Paris Cité.
She spoke about imaging science and presented the Plug and Play approach to address inverse problems related to image and video restoration.
Julie presented the limitations of neural networks in handling image restoration, even though they can be quite powerful in addressing simple inverse problems. The main limitation is their problem specific approach. This means that there is a need for retraining a neural network for each problem (for example, deblurring or denoising an image).
Then, Julie introduced us to a ‘Plug and Play’ approach that aims to address multiple inverse problems with only one Neural Network. They are a good solution for embeddable systems with limited memory.
For those interested, Julie and her team created a Github library (section: image restoration) with Jupyter notebooks.
If you want to know more, check out her slides below.
The concerns about Fairness in AI are growing as the new technology becomes increasingly pertinent in healthcare, job hiring, loan granting, and other domains.
Our second speaker, Ruta Binkyte, doctoral researcher at Inria Saclay-Ile-de-France, LIX, École Polytechnique, presented an interesting topic: identifying and mitigating bias in machine learning.
She started by presenting notable AI discrimination cases in face recognition and in healthcare, where bias has been demonstrated.
Bias can enter in any stage of the data science process but it mostly comes from the data. She raised two types of bias:
- Representation bias, where a group is underrepresented in a sample,
- Historical bias, where a historically disadvantaged group has lower occurrence of positive labels.
Ruta and her team worked particularly on these 2 biases and found that when representation bias is combined with historical bias, random data augmentation doesn’t always solve the problem and that adding samples with positive labels improves results.
She also spoke about more nuanced fairness notions and suggested strategies for overcoming or mitigating biases. Ruta introduced a framework created by her and her team to address bias challenges: Bayesian Bias Elimination (BaBE).
If you want to know more, check out her slides below.
Our last talk by Francesca Iannuzzi, Head of Data Science at Maison du Monde, was as interesting as it was fun.
She described a 3 year journey to develop a full sales forecasting product, highlighting the challenges that she and her team faced, which are common to many data science projects.
She specifically spoke about how they dealt with tight timelines, evolving demands and priorities changes. This resulted in changes to the team’s organisation, and the implementation of new processes and a change management strategy.
While they began with a manual process, they ended up today to a fully automated product combining Airflow, Google Cloud products Big Query and Vertex AI, and Qlik. Well done!
If you want to know more, check out her slides below.
If you do not want to miss our our events, you can:
🔗 follow us on Twitter, Meetup, and LinkedIn
📑 check our Google spreadsheet if you want to speak 📣, host 💙, help 🌠
📍join our Slack channel for information, discussions, and opportunities
📩 send an email to the Paris WiMLDS team to paris@wimlds.org
🎬 subscribe to our WiMLDS Paris Youtube channel
📸 follow the global WiMLDS on Instagram, LinkedIn, and Facebook
🔥 share your company or lab’s job offers for free on the global WiMLDS’ website.
