Opens profile photo
Follow
Click to Follow CChadebec
Clément Chadebec
@CChadebec
Ph.D @ INRIA. Working on Deep Generative Models and maintaining python packages democratizing them.
Joined April 2021

Clément Chadebec’s Tweets

Our experiment tracking tool helps you store experiment configs, track training, and compare the results in an easy and understandable way through a visual interface. We're now integrated with the Pythae library for easier ML monitoring!
5
We see the inverse of the covariance matrix in the posterior distribution as a local approximation of the Riemannian metric in the latent space at the embedding point. This allows to build a continuous smooth Riemannian metric on the whole latent space. 3/5
Image
1
6
Show this thread
We propose to interpret the latent space of VAEs as a Riemannian manifold and propose to sample uniformly from the learned manifold. We show that it improves the generated samples quality and can benefit more advanced VAE models such as AAE, VAMP-VAE, VAEGAN or IWAE. 2/5
Image
2
7
Show this thread
I sincerely hope you will enjoy these new features! 😀 This library is also open to new contributors so do not hesitate to reach out if you want to include a model, fix a potential bug 🐛 or would like to see any new feature!
3
Show this thread
4 - Finally, we validate the method on a medical imaging classification task on ADNI consisting in finding 3D MRIs of Alzheimer disease (AD) patients from cognitively normal (CN) ones. The model again showed useful to augment the data and allowed to improve a SOTA classifier.
Image
3
Show this thread
3 - Robustness to classifier complexity and dataset size. The model appears able to constantly increase the classification performances of the classifier regardless of it complexity. For the dataset size, the method still allow a performance gain even with more training samples.
Evolution of the accuracy of a benchmark DenseNet classifier according to the number of samples in the train set (i.e. the baseline) (left), the number of parameters of the Densenet (middle).
1
1
Show this thread
2 - Robustness across classifiers. In addition to a deep neural network (DenseNet), we also tested the method with 1) a Multi Layer Perceptron, 2) a SVM, 3) a kNN and 4) a Random Forest. In each case, the model revealed useful and allowed better classification performances.
Evolution of the accuracy of four benchmark classifiers on reduced balanced MNIST (left) and reduced unbalanced MNIST data sets (right). Stochastic classifiers are trained with five independent runs and we report the mean accuracy and standard deviation on the test set.
1
2
Show this thread
Then, we conduct a wide experimental study to show the usefulness of the method for Data Augmentation. 1 - Robustness to datasets. Only a fraction of training data is considered (~500) and augmented for each dataset and the classifier (DenseNet) is tested on the original test.
Data augmentation with a DenseNet model as benchmark. Mean accuracy and standard deviation across five independent runs are reported. The first three rows (Aug.) correspond to basic transformations (noise, crop, etc.). The test set is the one proposed in the entire original data set (e.g. ~1000 samples per class for MNIST) so that it provides statistically meaningful results and allows for a good assessment of the model's generalization power.
1
2
Show this thread
To enhance the generation capability of the model, we also propose a new generation scheme relying on the learned geometry of the latent space. It proves to be very efficient in the context of low sample size setting when compared to the prior or other post-training generations.
VAE sampling comparison. Using either the prior (left) or the proposed generation scheme (right).
1
1
Show this thread
This model has a the latent space modeled as a Riemannian manifold and combines both Riemannian metric learning and geometry-aware normalizing flows. It better captures the intrinsic geometry of the latent space and leads to convincing interpolations even with small datasets.
Geodesic interpolations under the learned metric in two different latent spaces.
1
5
Show this thread
🌸 BLOOM's intermediate checkpoints have already shown some very cool capabilities! What's great about BLOOM is that you can ask it to generate the rest of a text - and this even if it is not yet fully trained yet! 👶 🧵 A thread with some examples
Image
Quote Tweet
A milestone soon to be reached 🚀💫 Can't wait to see the capabilities and performance of this long-awaited checkpoint! What about you? Have you already prepared some prompts that you want to test? ✏️ twitter.com/BigScienceLLM/…
7
153
Show this thread
This library was also used to benchmark 19 generative autoencoders on 5 different tasks (image generation, reconstruction, classification, clustering and interpolation). All the models underwent a hyperparameter search, for 3 dataset and 2 autoencoding architectures.
Image
15
Show this thread
Pythae currently contains: 🚨21 unified implementations 🎲9 different sampling methods such as a Gaussian mixture model, normalizing flows (MAF, IAF) or PixelCNN 🧩2 pipelines (training and generation) ... making the code easy to use.
Image
1
12
Show this thread
Lots of great stuff coming up next week to accompany the release of the course! 🥳 1⃣ Release of part 2 of the course 2⃣ Two days of talks on November 15th-16th 3⃣ Event for you to put into practice NLP concepts and get a certificate 💫 Register here:
1
104