Jeux d'entrainement, de validation et de test

En apprentissage automatique, une tâche courante est l'étude et la construction d'algorithmes qui peuvent apprendre et faire des prédictions sur les données. De tels algorithmes fonctionnent en faisant des prédictions ou des décisions basées sur les données, en construisant un modèle mathématique à partir des données d'entrée. Ces données d'entrée utilisées pour construire le modèle sont généralement divisées en plusieurs jeux de données . En particulier, trois jeux de données sont couramment utilisés à différentes étapes de la création du modèle : les jeux d'apprentissage, de validation et de test. Le modèle est initialement ajusté sur un jeu de données d'apprentissage qui est un jeu d'exemples utilisés pour ajuster les paramètres (par exemple, les poids des connexions entre les neurones dans les réseaux de neurones artificiels ) du modèle. Le modèle (par exemple un classificateur naïf de Bayes ) est entraîné sur le jeu de données d'apprentissage à l'aide d'une méthode d'apprentissag

The EBRAINS Hodgkin-Huxley Neuron Builder: An online resource for building data-driven neuron models

Shailesh Appukuttan, Luca Leonardo Bologna, Jean-Denis Georges Emile Courcol, Genrich Ivaska, Carmen Alina Lupascu, Michele Migliore

In the last decades, brain modeling has been established as a fundamental tool for understanding neural mechanisms and information processing in individual cells and circuits at different scales of observation. Building data-driven brain models requires the availability of experimental data and analysis tools as well as neural simulation environments and, often, large scale computing facilities. All these components are rarely found in a comprehensive framework and usually require ad hoc programming. To address this, we developed the EBRAINS Hodgkin-Huxley Neuron Builder (HHNB), a web resource for building single cell neural models via the extraction of activity features from electrophysiological traces, the optimization of the model parameters via a genetic algorithm executed on high performance computing facilities and the simulation of the optimized model in an interactive framework. Thanks to its inherent characteristics, the HHNB facilitates the data-driven model building workflow and its reproducibility, hence fostering a collaborative approach to brain modeling.

FRONTIERS MEDIA SA2022

Robustness via Cross-Domain Ensembles

Oguzhan Fatih Kar, Amir Roshan Zamir, Shuqing Teresa Yeo

We present a method for making neural network predictions robust to shifts from the training data distribution. The proposed method is based on making predictions via a diverse set of cues (called 'middle domains') and ensembling them into one strong prediction. The premise of the idea is that predictions made via different cues respond differently to a distribution shift, hence one should be able to merge them into one robust final prediction. We perform the merging in a straightforward but principled manner based on the uncertainty associated with each prediction. The evaluations are performed using multiple tasks and datasets (Taskonomy, Replica, ImageNet, CIFAR) under a wide range of adversarial and non-adversarial distribution shifts which demonstrate the proposed method is considerably more robust than its standard learning counterpart, conventional deep ensembles, and several other baselines.

IEEE2021

Relative stability toward diffeomorphisms indicates performance in deep nets

Alessandro Favero, Mario Geiger, Leonardo Petrini, Matthieu Wyart

Understanding why deep nets can classify data in large dimensions remains a challenge. It has been proposed that they do so by becoming stable to diffeomorphisms, yet existing empirical measurements support that it is often not the case. We revisit this question by defining a maximum-entropy distribution on diffeomorphisms, that allows to study typical diffeomorphisms of a given norm. We confirm that stability toward diffeomorphisms does not strongly correlate to performance on benchmark data sets of images. By contrast, we find that the {\it stability toward diffeomorphisms relative to that of generic transformations}

R_f

correlates remarkably with the test error

\epsilon_t

. It is of order unity at initialization but decreases by several decades during training for state-of-the-art architectures. For CIFAR10 and 15 known architectures we find

\epsilon_t\approx 0.2\sqrt{R_f}

, suggesting that obtaining a small

R_f

is important to achieve good performance. We study how

R_f

depends on the size of the training set and compare it to a simple model of invariant learning.

2021