Emma Sarfati

ML Engineer @Criteo R&D | PhD in Deep Learning @TelecomParis @GuerbetGroup

[Emma Sarfati @EmmaSarfati_]

This work led to:
📝 3 conference papers (ISBI’23, MICCAI’23, ISBI’25)
📄 1 abstract at ECR’24
📑 1 journal submission under review

📕 Thesis manuscript: https://t.co/kSwGRwWcmh
🔗 Google Scholar: https://t.co/UY17NRIZt5

#MedicalImaging #AI #DeepLearning

3/3

[Emma Sarfati @EmmaSarfati_]

🚀 Thrilled to share that I defended my PhD in ML & Medical Imaging (Télécom Paris × Guerbet Research) on June 19, 2025!

Title: “Exploiting Heterogeneous Labels in Deep Learning for Medical Image Analysis: Application to the Automated Diagnosis of Liver Diseases”

1/3

[Emma Sarfati @EmmaSarfati_]

In this thesis, I studied how to improve classification models under limited supervision using weakly-supervised and contrastive learning approaches, leveraging partial labels and structured metadata for cirrhosis and HCC diagnosis.

#PhDDefense #ML #MedicalImaging

2/3

[Ryan @RyanBoustany_]

🚀 Excited to share our new preprint with @jerome_bolte, E. Pauwels & A. Purica: "A second-order-like optimizer with adaptive gradient scaling for deep learning."

Check out the paper: https://t.co/JiYmwcsqPw
GitHub: https://t.co/uIV9Un7XmR

(1/3)

[Ryan @RyanBoustany_]

Happy to share my new preprint on the numerical reliability of nonsmooth autodiff. Using a MaxPool case study, I tried to study the behavior of nonsmooth AD across different precision levels (16, 32, and 64 bits).

https://t.co/mB6cbDFZOA

[Emma Sarfati @EmmaSarfati_]

Code is now out here: https://t.co/5x1uFV2ALs

[Emma Sarfati @EmmaSarfati_]

New accepted paper at #MICCAI2023! In this paper, we leverage continuous and discrete metadata for contrastive learning via a new generic kernel-based loss function. We apply our work to cirrhosis classification using pretraining methods.

Paper: https://t.co/eSwZ3ZkBiq

[Emma Sarfati @EmmaSarfati_]

@kornia_foss Thanks! @kornia_foss 🙏🏼 very great library, guys if you are limited in RAM use it for data augmentation it changes everything!

[Kornia @kornia_foss]

Code is (not yet) here: https://t.co/FPlXzcNBF5

[Kornia @kornia_foss]

Weakly-supervised positional contrastive learning: application to cirrhosis classification
@EmmaSarfati_, Alexandre Bône, Marc-Michel Rohé, Pietro Gori, Isabelle Bloch

tl;dr: use volumetric nature of medical images for learning.
#kornia -> data aug.
https://t.co/0aEuMSUrpc

[Emma Sarfati @EmmaSarfati_]

@CopaAirlines Are you kidding me? Your agent proposed me a voucher for your company, but I live in France and you have literally no flight coming or going to Europe!!!!!!!!! I need a compensation, it is a scandal! You can be sure that I won’t recommand your company to anyone

[Emma Sarfati @EmmaSarfati_]

If one day you plan to go on a trip with @CopaAirlines, please don’t. Except if you want to wait 10 hours on the ground of an AirPort during a connection, without having any coffee and any explanation of the situation. And of course if you are European, you will have to accept

[Emma Sarfati @EmmaSarfati_]

@CopaAirlines Hello, your account is automated by prívate message, I cannot send something I need to click on options that redirect me to some links. How can I do please?

[Emma Sarfati @EmmaSarfati_]

that your compensation will be a voucher for a flight with Copa, which literally has no flight coming or going to Europe. Some pictures of the post chaotic experiment of my life two weeks ago when going to Cartagena for #isbi. You are scandalous @CopaAirlines

[Ryan @RyanBoustany_]

We also show that conservative gradients have computational properties similar to smooth derivatives, which are much more favorable than those of alternative nonsmooth oracles such as subgradients or directional derivatives. (2/2)

[Ryan @RyanBoustany_]

Pleased to share our new preprint with J.Bolte, E.Pauwels and B.Pesquet-Popescu! Using conservative gradients, we establish a “nonsmooth cheap gradient principle” for backpropagation encompassing most concrete applications. (1/2)
Preprint : https://t.co/Y9H0TOXnKY