Cette journée est organisée par le Laboratoire d’Informatique de Paris Nord LIPN a pour but de présenter des travaux en intelligence artificielle, soit réalisés par des chercheurs du LIPN soit réalisés par des chercheurs extérieurs aussi bien académiques qu’industriels.
Cette journée aura lieu dans l’amphithéâtre Euler de l’institut Galilée.
09h15 Ouverture de la journée par Frédérique Bassino, directrice du LIPN
Matinée Apprentissage profond
Designing deep architectures for Visual Question Answering
Mixing text and image in deep net representations has been extensively studied in recent years. One of the most popular tasks in Computer Vision on this topic is Visual Question Answering (VQA). I will introduce this multimodal VQA task, which aims at answering a question about an image. To solve this problem, not only visual and textual deep nets models are required, but also high level interactions between these two modalities. Besides, the model must have the ability to fully understand the visual scene, focusing on the relevant visual regions and discarding the useless information regarding the question. Two publications AAAI 2019 and CVPR 2019 will illustrate this talk.
Learning Deep Representations Through Stacked Prototype Based Models
Hierarchical structures are known since decades for their outstanding properties that make them ideal for representing data and has been suggested as a particularly important method for organizing concepts. Deep unsupervised networks are useful approaches to learn a synthetic representation of the underlying data distribution. In this work we explore an original strategy for building deep networks, based on stacking layers of Self-Organizing Maps (SOM) with finite weights. The first layer receives the observations and compute a probability of membership for each observation and each neuron. Each layer after the first receive as input the probabilities computed in the previous layer and produce new membership values. The output of the network is a probabilistic representation of the input data. In addition, we present a Multilayer Nonnegative Matrix Factorization approach. We explicitly find the dependence function linking different layers in the process and investigate the relative role of each layer in the hierarchy.
10h45-11h15 Pause Café
Différentiation automatique dans les programmes fonctionnels
Spoken Language Understanding on the Edge
I will present the machine learning architecture of the Snips Voice Platform, a software solution to perform Spoken Language Understanding on microprocessors typical of IoT devices. The embedded inference is fast and accurate while enforcing privacy by design, as no personal user data is ever collected. I will detail our approach to training high-performance Machine Learning models that are small enough to run in real-time on small devices. Additionally, I will describe a data generation procedure that provides sufficient, high-quality training data without compromising user privacy.
Part of my talk will be based on this publication: https://arxiv.org/abs/1810.12735
Representation Learning and Dynamic Programming for Transition-based Parsing
Transition-based, or shift-reduce, dependency parsing is a core task in Natural Language Processing and is traditionally implemented as a beam-search procedure. However some recent work showed that global inference can be performed efficiently with dynamic programming. We first show how a solution in transition-based parsing can be represented as a pair consisting of a derivation graph and a derived dependency tree allowing the scoring function to be expressed naturally over these 2 structures. We then propose an alternative approach to global inference where derivation steps are represented as dense vectors based on the number and type of steps in a derivation. With this abstraction we design a neural architecture based on non-local networks and a self-attention mechanism, to learn these representations while maintaining the possibility for exact decoding. This new representation achieves state-of-the-art performance on main test sets.
Après-midi Fouille de Graphe
Onto Model-Based Exceptional Pattern Mining on Complex Interaction Networks
The detection of anomalies and exceptional patterns in complex interaction networks is a prominent research direction in complexity and network science. Typically two questions need to be addressed: (1) What is an exceptional/anomalous pattern ? (2) How do we identify that? This talk presents model-based approaches and methods for addressing and formalizing these issues in the context of complex interaction networks, and exemplifies promising directions for its implementation.
A short tutorial on community detection : from plain to feature-rich complex networks
15h30-16h00 Pause café
Mining Augmented Graphs: Some contributions and a research agenda
Graphs are a powerful mathematical abstraction that enables to depict many real world phenomena. Vertices describe entities and edges identify relations between entities. Such graphs are often augmented with additional pieces of information. For instance, the vertices or the edges are enriched with attributes describing them and are called vertex (respectively edge) attributed graphs. Graphs can also be dynamic, i.e., the structure and the values of vertex attributes may evolve through time. The discovery of patterns in such graphs may provide actionable insights and boost the user knowledge. In this talk, I will discuss the different pattern domains for augmented graphs we contributed to define. This includes the discovery of exceptional attributed subgraphs in edge or vertex attributed graphs. Then, I will discuss how to find patterns of higher interest by taking into account the domain knowledge, user feedback and user’s prior knowledge through different examples.
Bi-Pattern Mining of Attributed Networks
To apply closed pattern mining to attributed two-mode networks requires two conditions. First, as there are two kinds of vertices, each described with a proper attribute set, we have to consider patterns made of two components that we call bi-patterns. The occurrences of such a bi-pattern forms then an extension made of a pair of vertex subsets. Second, Formal Concept Analysis and Closed Pattern Mining were recently applied to networks by reducing the extensions of pattern extensions to their cores, according to some core definition. To apply this methodology to two-mode networks, we need to consi der two mode cores and define accordingly abstract closed bi-patterns. We give in this article a general framework to define closed bi-pattern mining. We also show that the same methodology applies to cores of directed and undirected networks in which each vertex subset is associated to a specific role. We illustrate the methodology on a two-mode network of epistemological data, on a directed advice network of lawyers, and on two undirected networks.
Mining Scholarly Data for Fine-Grained Knowledge Graph Construction
Graphs are frequently used to express quantitative and qualitative information through relatively simple patterns. Projects like DBPedia, Google Knowledge Graph, YAGO, BabelNet, among others, harvest entities and relations from Wikipedia textual resources and organize the mined information into graphs. Data sources where information are harvested determine the quality of results since methods exploited for lexicon analysis strongly depend on the syntax and the semantics of the underlying contents. Previous projects addressed the problem from a generic perspective, without considering the type of information that data contain. In this talk, we present a workflow to build a fine-grained Knowledge Graph from Semantic Web Scholarly Data, and propose empirical solutions for refining its entities and relations.