Search

The deliverable primarily covers the initial implementation of the Policies and Explainability and Interpretation Framework V1. It describes and demonstrates the tools and algorithms developed during this implementation phase. The document is structured into several chapters, starting with an introduction to the project and the deliverable. The subsequent chapters delve into the background, state of the art, solution design, implementation details, and finally, the conclusions and future works. Overall, this deliverable lays the groundwork for the project's AI models and tools, setting the stage for further development and refinement in subsequent phases.

International audience; Jacovi et al. in 2018 proposed a method for the explanation of Convolutional Neural Networks designed for text classification (Text-CNN). The problem with this method is that it only applies to Text-CNN with no hidden layer in the fully connected part. We will therefore find ourselves limited in terms of accuracy if we want to use this method. In addition, the LRP (Layer-wise Relevance Propagation) method makes it possible to compute the contributions of inputs of neural networks with multiple hidden layers, but cannot be applied directly to Text-CNNs because unlike images where each pixel has a meaning taken alone, each component of a vector representation (word embedding) of a word taken alone has no particular meaning. Therefore, we propose in this article to extend the method of Jacovi et al. by using the LRP method in order to be able to explain Text-CNN whose fully connected layer is deep. The qualitative evaluation suggests that the explanations provided by the proposed method are consistent on text classification problems such as Question-Answering (QA) and Sentiment Analysis. Moreover, the quantitative evaluation shows that the explanations provided are 100% faithful to the model.; Jacovi et al. en 2018 ont proposé une méthode pour l'explication des Réseaux de Neurones Convolutifs conçus pour la classification des textes (Text-CNN). Le problème avec cette méthode est qu'elle ne s'applique que sur les Text-CNN qui n'ont pas de couche cachée dans la partie densément connectée. On se retrouvera donc limité en terme de performance si on veut utiliser cette méthode d'explication. Par ailleurs, la méthode LRP (Layer-wise Relevance Propagation) permet de calculer les contributions des caractéristiques en entrée des réseaux de neurones avec un nombre quelconque de couches cachées, mais ne peut s'appliquer directement sur les Text-CNN car contrairement aux images où chaque pixel a un sens pris seul, chaque composante d'une représentation vectorielle (word embedding) d'un mot prise seule n'a pas de signification particulière. C'est ainsi que nous proposons dans cet article, d'étendre la méthode de Jacovi et al. en mettant à contribution la méthode LRP afin de pouvoir expliquer les Text-CNN dont la partie densément connectée est profonde (possède plusieurs couches). L'évaluation qualitative laisse percevoir que les explications fournies par la méthode proposée sont cohérentes sur les problèmes de classification des textes comme le Question-Answering (QA) ainsi que l'analyse des sentiments. Par ailleurs, l'évaluation quantitative montre que les explications fournies sont fidèles à 100% au modèle.

This paper aims at bringing some light and understanding to the field of deep learning for dynamic point cloud processing. Specifically, we focus on the hierarchical features learning aspect, with the ultimate goal of understanding which features are learned at the different stages of the process and what their meaning is. Last, we bring clarity on how hierarchical components of the network affect the learned features and their importance for a successful learning model. This study is conducted for point cloud prediction tasks, useful for predicting coding applications.

International audience; Attention maps in neural models for NLP are appealing to explain the decision made by a model, hopefully emphasizing words that justify the decision. While many empirical studies hint that attention maps can provide such justification from the analysis of sound examples, only a few assess the plausibility of explanations based on attention maps, i.e., the usefulness of attention maps for humans to understand the decision. These studies furthermore focus on text classification. In this paper, we report on a preliminary assessment of attention maps in a sentence comparison task, namely natural language inference. We compare the cross-attention weights between two RNN encoders with human-based and heuristic-based annotations on the eSNLI corpus. We show that the heuristic reasonably correlates with human annotations and can thus facilitate evaluation of plausible explanations in sentence comparison tasks. Raw attention weights however remain only loosely related to a plausible explanation.