Chen Liu

AP @CityUHongKong, PhD MS @EPFL_en, UG @Tsinghua_Uni, machine learning, optimization, robustness, privacy.

[Chen Liu @liuchen1993]

(5/5) We have one workshop paper on GenAI in Finance https://t.co/OFd9EEhD6z which uses LLMs for genetical programming for effective and efficient alpha mining. The assets selected by our alpha factors have better long-term returns than traditional ML in US/HK/CN market.

[Chen Liu @liuchen1993]

(4/5) Our work in NeurIPS about data selection for LLM fine-tuning https://t.co/81VT9czQ1Q designs how to select data based on robustness for efficient instruction tuning. On several datasets, we can compress the training set to its 30% without performance loss after finetuning.

[Chen Liu @liuchen1993]

(3/5) Our work in NeurIPS about fast sparse adversarial training https://t.co/axOIf6ZAl5 studies the loss landscape of fast AT against sparse perturbations. We propose smoothing techniques to improve the algorithm performance and stability.

[Chen Liu @liuchen1993]

(2/5) Our work in NeurIPS Dualoptim https://t.co/9PyHS77UzS greatly improves the stability of machine unlearning with a simple trick: separate optimizers to the forget and the retain. It is effective for popular MU methods and models: classification/diffusion models/LLMs.

[Chen Liu @liuchen1993]

(1/5) Recent update: I will go to NeurIPS 2025 and present 3 conf papers and 1 workshop paper. I look forward to seeing you in San Diego : )

[Chen Liu @liuchen1993]

It was the first work of my first PhD student, despite not his first published one.

[Chen Liu @liuchen1993]

Our work about mitigating architecture overfitting in dataset distillation is updated on https://t.co/cVfIsc2NxT and will be published in IEEE TNNLS. It discusses different techniques in architecture, algorithm and optimizer to solve this problem by comprehensive case studies.

[Chen Liu @liuchen1993]

@WEIMA91 Thank you 😄

[Chen Liu @liuchen1993]

Our JMLR paper is online (https://t.co/jKLIrpd7RI) after almost three years.. it is about theoretical analyses of overfitting issue in adversarial training. (I finally found my account passwords to tweet..

[Chen Liu @liuchen1993]

“You have agreed to review up to -1 papers”

[Chen Liu @liuchen1993]

Change the descent geometry and got rid of coordinate descent can greatly mitigate this issue and make more efficient training possible. Cmp with existing methods, our simple algorithm has no more hyper params, better efficiency, almost no memory overhead and better performance.

[Chen Liu @liuchen1993]

At #ICML, we will present a work about how to efficiently and stably obtain robust deep neural net for l1 bounded perturbations, a different story from l2 and linf cases. paper: https://t.co/7ICABcUtTe. Code: https://t.co/XdN9Ac75hm. Look forward to the events in Hawaii

[Chen Liu @liuchen1993]

key take away: coordinate descent used in existing methods for l1 robustness tend to generate sparse perturbations and leads to more frequent catastrophic overfitting in l1 case, even when training against multi step attacks.

[Chen Liu @liuchen1993]

@maksym_andr congrats, see you at Hawaii

[Chen Liu @liuchen1993]

@deblinaforAI congrats!🎉

[Chen Liu @liuchen1993]

Source code on https://t.co/QqHOcu3LP6; paper on https://t.co/H5f5LLQprP

[Chen Liu @liuchen1993]

Poster 124, this afternoon 😆

[Chen Liu @liuchen1993]

Happy to see this work will be published in #NeurIPS2022 . An updated version and the source code will come soon!

[Ola Svensson @AlgoSvensson]

Amazing event at @EPFL on Friday!

Zoom link: https://t.co/T3cx1iioWK

[Maksym Andriushchenko @maksym_andr]

I'll give a talk about our ICML'22 paper https://t.co/F63P1KkdCw tmrw (3pm CET) at the ELLIS Mathematics of Deep Learning reading group.

The zoom link is available here: https://t.co/X2tjSJwm04. Feel free to drop by if you want to chat a bit about sharpness and related topics :)

[Alhussein Fawzi @AlhusseinFawzi]

Excited to share our paper in @Nature: We revisit the 50+ year-old maths problem with AI: how efficiently can we multiply two matrices? Surprisingly, the answer is still not known - even for 3x3 matrices! With AI, we discover many new efficient and exact algorithms. 1/13