Ryan

PhD candidate in ML - @thalesgroup & @TSEinfo & @ANITI_Toulouse.

[Ryan @RyanBoustany_]

Accepted at @NeurIPSConf 2025 😎

[Ryan @RyanBoustany_]

📢New preprint
“When majority rules, minority loses: bias amplification of gradient descent”

We often blame biased data but training also amplifies biases. Our paper explores how ML algorithms favor stereotypes at the expense of minority groups.

➡️https://t.co/vFnDoQobqJ

(1/3)

[Paul Couairon @PaulCouairon]

🚀Thrilled to introduce JAFAR—a lightweight, flexible, plug-and-play module that upsamples features from any Foundation Vision Encoder to any desired output resolution (1/n)

Paper : https://t.co/le4pF8rVXH
Project Page: https://t.co/rLW3jbin3O
Github: https://t.co/1AL7ElibHf

[UEFA Champions League @ChampionsLeague]

Legend.

#UCLfinal

[Ryan @RyanBoustany_]

💡 How much more training is needed to move from a biased to a fair predictor ?
Empirically: up to 400% extra training.

Joint work with @FBachoc, @jerome_bolte, @JM_Loubes.

(3/3)

[Ryan @RyanBoustany_]

📢New preprint
“When majority rules, minority loses: bias amplification of gradient descent”

We often blame biased data but training also amplifies biases. Our paper explores how ML algorithms favor stereotypes at the expense of minority groups.

➡️https://t.co/vFnDoQobqJ

(1/3)

[Ryan @RyanBoustany_]

- The majority loss dominates the landscape: minimizing the loss is almost always equivalent to minimizing the majority part.

- We introduce a stereotype gap: the distance between fair and majority-optimal critical points.

- Debiasing requires extra training.

(2/3)

[Guillaume Astruc @g_astruc]

We've added new experiments demonstrating robust generalization capabilities! Notably, AnySat shows strong performance on HLS Burn Scars - a sensor never seen during pretraining! 🔥🛰️
Check it out:
📄 Paper: https://t.co/PIKB2w6K1Z
🌐 Project: https://t.co/Qa59zWSJZe

[Guillaume Astruc @g_astruc]

🤔 What if embedding multimodal EO data was as easy as using a ResNet on images?
Introducing AnySat: one model for any resolution (0.2m–250m), scale (0.3–2600 hectares), and modalities (choose from 11 sensors & time series)!
Try it with just a few lines of code:

[Jérôme Bolte @jerome_bolte]

Recruiting Post-Doctoral students in Machine-Learning, Optimization or Regulation in AI at TSE or Paris. Write to [email protected]

[Ryan @RyanBoustany_]

INNAprop outperforms or matches AdamW in training speed and accuracy on CIFAR-10, ImageNet, and language modeling (GPT-2).

Here is an example of a perplexity test with GPT-2 and E2E Dataset using LoRA finetuning.

(3/3)

[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_]

Our novel optimizer INNAprop combines Hessian driven damping with the RMSprop adaptive gradient scaling.

It leverages second-order information and rescaling while keeping the computational and memory requirements of AdamW.

(2/3)

[cocktail peanut @cocktailpeanut]

OpenAI preparing to drop their new model

[Samuel Vaiter @vaiter]

📢Preprint alert📢

**Geometric and computational hardness of bilevel programming**

w/ @jerome_bolte, Tùng Lê & Edouard Pauwels

We study how difficult it may be to solve bilevel optimization beyond strongly convex inner problems

https://t.co/i6dghCCaxo

[Guillaume Astruc @g_astruc]

Introducing OmniSat: Self-Supervised Modality Fusion for Earth Observation, accepted at #ECCV2024 with @NicaoGr, Clément Mallet and @captnloic 🛰️🌍
OmniSat exploits the spatial alignment between modalities to learn expressive multimodal representations without labels.

[Guillaume Astruc @g_astruc]

I will be at #CVPR2024 this week presenting:

📅 Fri AM, nº 233
🗺️OpenStreetView-5M: The Many Roads to Global Visual Geolocation:
https://t.co/PADwIzft46

Feel free to reach out if you want to discuss geolocation, remote sensing, self-supervised learning and cross modal learning!

[Ryan @RyanBoustany_]

Published in @TmlrOrg 😎

https://t.co/ZaElGN0YM5

[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

[Loic Landrieu @loiclandrieu]

Can you guess where these photos were taken? @geoguessr players have taken this skill to the extreme, but how good can an AI perform? Introducing OpenStreetView-5M 🌍, the first open-access and global-scale dataset of street view images.
🔗 Links and 🤖 demo below 👇
#CVPR2024

[Michael Arbel @MichaelArbel]

📢 *PhD opening* at
@inria_grenoble
! Edouard Pauwels, @vaiter and myself are looking for a student to work with us on learning theory for bilevel optimization, in particular, the implicit bias in bilevel optimization.

If interested, please reach out!

[Samuel Vaiter @vaiter]

I have potentially an opening for a postdoc position (24 months, Inria contract) to work at @LJADnice @Univ_CotedAzur on mathematical guarantees for automatic differentiation. DM/email me if interested