Re-writing a non-convex optimization as a difference of convex (DC) leads to a simple iterative optimization method. Sometimes referred to as "convex-concave procedure" in the ML literature. https://t.co/b9kWb2km25 https://t.co/0NxVFMxxxG

It took us 8 years to reach this. Open Source Software is a marathon not a sprint. If you want it to make an impact you need to update and maintain it. This takes a lot of (personal) time, it is not always easy but you will meet good OSS friends and it is totally worth it.

From the citations, we can see that POT is used on biomedical data, machine learning applications such as data alignment, graphs and generative modeling but also physics. I was also very surprised to see it used/cited by the ATLAS experiment at CERN.

POT now have 75 contributors who made it what it is today 🫶, has been downloaded nearly 3M times and is part of Debian/Ubuntu repositories (ultimate geek bucket list for me). https://t.co/fnFNk1HRpX

With some help (and terrifying but so nice code reviews) from @agramfort we made it a proper Python library with a documentation and an example gallery. https://t.co/LTPHkuj3kh

For those wondering we started POT with @nicolas_courty with a Sinkhorn solver for entropic OT and a cython wrapper adapted from a Matlab mex file (remember those?) using the C++ solver from @nbonneel for exact OT.

Today something crazy happened. POT has reached 1000 citations (total) 🤩🚀. Very proud to be part of a scientific community that acknowledges open source research software. Please continue to use, cite and contribute to POT ! Small🧵below for those interested

I will be with Paul at NeurIPS@Paris next week and Paul will go present the paper at the main conference in December. Come and see us!

The performance of our Any2Graph method on graph datasets with different modalities is very good and robust to the maximum graph size parameter M.

Finally we propose an asymmetric Optimal Transport Loss (PM-FGW) that measure the quality of the prediction using both the continuous prediction and the discrete target object especially the binary padding mask h

Next we use a transformer encoder/decoder architecture to predict from any input a continuous version of the adjacency matrix and masks (binary classification) and the node features.

First we pad all graphs to a maximum size M. We add to the Adjacency matrix a padding mask h that define which nodes are activated

We have an accepted paper at NeurIPS 2024 about supervised graph prediction with transformers and optimal transport loss with Paul Krzakala, Junjie Yang, Florence d’Alché-Buc, Charlotte Laclau and Matthieu Labeau. Details in 🧵 https://t.co/jP35r8rIhn

Those demos use @pygame_org , @opencvlibrary and of course @scikit_learn

2D Dictionary learning: dynamic demo where user clicks to add 2D sample with visualization of the dictionary atoms for PCA, ICA, NMF, Sparse DL and KMeans

2D filtering Demo: visualization of the 2D filtering of the webcam and its spectrum in real time

2D FFT Demo: visualization of the 2D FFT absolute value of the centered spectrum in linear and log scale in real time.

Just updated my Python real time demos. Added a visualization of 2D FFT of webcam, of 2D filters of webcam, and a dynamic example of dictionary learning with atoms. Love to use those in my courses 🧑‍🔬. (see examples in thread)

This is a very interesting letter (in french) with concrete recommendation for more fairness and diversity in academic recruitment. We should all read this. Especially those who write recommendation letters and participate to recruitment committees

Python Optimal Transport (POT) 0.9.5 released: new solvers for Gaussian Mixture Model OT, unbalanced OT, semi-relaxed (F)GW barycenters, unbalanced FGW and COOT, partial GW. more details in 🧵1/7 https://t.co/c0TCTHQTnR