robocrunch
Olivier Grisel
@ogrisel
Engineer at @Inria, scikit-learn developer supported by https://t.co/6xVeamXYPa. Tweets about Python and Machine Learning / Deep Learning.
Tweets by Olivier Grisel
Enlightening tutorial-style paper on probabilistic model evaluation that discusses calibration and proper score functions such as the log loss, the Brier score, the Tweedie deviance for expectile estimators and the pinball loss for quantile estimators. https://t.co/394q3dOwXP
Shared by
Olivier Grisel
at
5/14/2022
I have been testing this branch with scikit-learn and there are still some open issues, e.g.: https://t.co/r8RkiORgiL but @colesbury is very fast at fixing them, sometimes in third-party libraries like numpy and Cython and this work will hopefully be upstreamed.
Shared by
Olivier Grisel
at
5/12/2022
For information, scikit-learn now has a nightly CI run on using the nogil fork: https://t.co/iVmTwXuG2R and it's green: https://t.co/pj3lrUEFmP the goal is to make sure scikit-learn stays compatible with the nogil fork.
Shared by
Olivier Grisel
at
5/12/2022
Interesting summary of the discussion of @colesbury's nogil work at the Python language summit: https://t.co/AQEZeDKL8a As a @scikit_learn, joblib and loky maintainer, I am very excited about this work.
Shared by
Olivier Grisel
at
5/12/2022
TIL that selecting node splits of a classification tree using the Shannon entropy criterion as a measure of leaf purity is strictly equivalent to (greedily) minimizing the log loss (or cross-entropy) of the predictions of the tree on the training set. https://t.co/Gm6IqWI5Yg
Shared by
Olivier Grisel
at
4/11/2022
Our team at Inria Saclay is looking for a software engineer interested in developing a plugin system to allow for efficient GPU computing kernels for popular machine learning algorithms in scikit-learn (nearest neighbors, k-means, T-SNE...): https://t.co/8XyzWfDxVp #job
Shared by
Olivier Grisel
at
3/24/2022
A very simple reparametrization + init scheme that makes it possible to train very deep transformers that generalize well on large multiway machine translation tasks. https://t.co/aqR0Z9cJrO
Shared by
Olivier Grisel
at
3/2/2022
I find the hierarchical JEPA approach very interesting and promising but you should acknowledge that what most readers will remember from this post is "Y. LeCun invented World Models" bc the post lacks a discussion on related works as would be expected in academic communication.
Shared by
Olivier Grisel
at
2/24/2022
Thanks to @SebConort's team at BNP Paribas Cardiff for supporting our work on the development and maintenance of scikit-learn via the @sklearn_inria consortium.
Shared by
Olivier Grisel
at
11/18/2021
Explaining machine learning models with SHAP and SAGE https://t.co/qPoLmWDY4Z SAGE is Shapley values for a loss function of a predictive model. This is similar to Breiman's permutation importance while correcting the pbm of low importance assigned to features w/ good proxies.
Shared by
Olivier Grisel
at
11/4/2021
and you can change the number of threads of a running Python program using the context manager pattern:
Shared by
Olivier Grisel
at
10/1/2021
threadpoolctl 3.0.0 is out: this is a Python utility to introspect & control the number of threads used by OpenMP parallel loops and BLAS linear algebra kernels, typically used in computational libraries (eg NumPy, SciPy and machine learning libraries). https://t.co/TqSb2LXRaJ
Shared by
Olivier Grisel
at
10/1/2021
We recently discovered that AdaBoost's training error can sometimes increase with more trees. Apparently an easy fix is to reset the weights whenever the beta error ratio degrades. Were this problem and its solution known in the literature? https://t.co/pgxwmFtii8
Shared by
Olivier Grisel
at
7/4/2021
Interesting use case for advanced second order optimizers like Shampoo: distillation with function matching where overfitting never happens:
Shared by
Olivier Grisel
at
6/10/2021
Interesting empirical study of the geometry of the loss landscape of Vision Transformers and MLP-Mixers and study of the critical impact of Sharpness Aware Minimization (SAM) for those architectures.
Shared by
Olivier Grisel
at
6/4/2021