Learning representations for Information Retrieval

La recherche d'informations s'intéresse, entre autres, à répondre à des questions comme: est-ce qu'un document est pertinent à une requête ? Est-ce que deux requêtes ou deux documents sont similaires ? Comment la similarité entre deux requêtes ou documents peut être utilisée pour améliorer l'estimation de la pertinence ? Pour donner réponse à ces questions, il est nécessaire d'associer chaque document et requête à des représentations interprétables par ordinateur. Une fois ces représentations estimées, la similarité peut correspondre, par exemple, à une distance ou une divergence qui opère dans l'espace de représentation. On admet généralement que la qualité d'une représentation a un impact direct sur l'erreur d'estimation par rapport à la vraie pertinence, jugée par un humain. Estimer de bonnes représentations des documents et des requêtes a longtemps été un problème central de la recherche d'informations. Le but de cette thèse est de proposer des nouvelles méthodes pour estimer les représentations des documents et des requêtes, la relation de pertinence entre eux et ainsi modestement avancer l'état de l'art du domaine. Nous présentons quatre articles publiés dans des conférences internationales et un article publié dans un forum d'évaluation. Les deux premiers articles concernent des méthodes qui créent l'espace de représentation selon une connaissance à priori sur les caractéristiques qui sont importantes pour la tâche à accomplir. Ceux-ci nous amènent à présenter un nouveau modèle de recherche d'informations qui diffère des modèles existants sur le plan théorique et de l'efficacité expérimentale. Les deux derniers articles marquent un changement fondamental dans l'approche de construction des représentations. Ils bénéficient notamment de l'intérêt de recherche dont les techniques d'apprentissage profond par réseaux de neurones, ou deep learning, ont fait récemment l'objet. Ces modèles d'apprentissage élicitent automatiquement les caractéristiques importantes pour la tâche demandée à partir d'une quantité importante de données. Nous nous intéressons à la modélisation des relations sémantiques entre documents et requêtes ainsi qu'entre deux ou plusieurs requêtes. Ces derniers articles marquent les premières applications de l'apprentissage de représentations par réseaux de neurones à la recherche d'informations. Les modèles proposés ont aussi produit une performance améliorée sur des collections de test standard. Nos travaux nous mènent à la conclusion générale suivante: la performance en recherche d'informations pourrait drastiquement être améliorée en se basant sur les approches d'apprentissage de représentations.

Looking for redemption in a globalised North:representations of the Arctic in Judith Hermann’s Short Stories Kaltblau (Cold –Blue) and Die Liebe zu Ari Oskarsson (Love for Ari Oskarsson)

This paper explores the literary representation of Iceland and Norway in two short stories by contemporary German writer Judith Hermann. It analyses both the depiction of these countries as part of the globalised western world and the redemptive power they are tentatively ascribed by the author. Continuing a long German tradition of looking at Scandinavia from an almost colonial perspective, Hermann on the one hand presents these northern countries as a mere extension of central Europe, largely devoid of distinguishing national characteristics. At the same time she makes reference to the topos of the north as a vast and empty space and highlights both the specific arctic nature of the environment and the effect it has on her urban characters, who find themselves on a search for meaning and orientation in a postmodern fragmented world. Despite Hermann's overall sceptical attitude towards her characters' quest for happiness, these northern locations ultimately appear as potential places of self-realisation and enlightenment.

The Riesz transform and simultaneous representations of phase, energy and orientation in spatial vision

To represent the local orientation and energy of a 1-D image signal, many models of early visual processing employ bandpass quadrature filters, formed by combining the original signal with its Hilbert transform. However, representations capable of estimating an image signal's 2-D phase have been largely ignored. Here, we consider 2-D phase representations using a method based upon the Riesz transform. For spatial images there exist two Riesz transformed signals and one original signal from which orientation, phase and energy may be represented as a vector in 3-D signal space. We show that these image properties may be represented by a Singular Value Decomposition (SVD) of the higher-order derivatives of the original and the Riesz transformed signals. We further show that the expected responses of even and odd symmetric filters from the Riesz transform may be represented by a single signal autocorrelation function, which is beneficial in simplifying Bayesian computations for spatial orientation. Importantly, the Riesz transform allows one to weight linearly across orientation using both symmetric and asymmetric filters to account for some perceptual phase distortions observed in image signals - notably one's perception of edge structure within plaid patterns whose component gratings are either equal or unequal in contrast. Finally, exploiting the benefits that arise from the Riesz definition of local energy as a scalar quantity, we demonstrate the utility of Riesz signal representations in estimating the spatial orientation of second-order image signals. We conclude that the Riesz transform may be employed as a general tool for 2-D visual pattern recognition by its virtue of representing phase, orientation and energy as orthogonal signal quantities.

“Seismic cultural change?”: British media representations of same-sex ‘marriage’

The legal recognition of same-sex relationships is a contested terrain that has been hotly debated by feminists. This article provides a social constructionist analysis of the UK newspaper media coverage around the time of the introduction of the Civil Partnership Act (2004). In examining the 348 national newspaper coverage over a three month period (November 2005–January 2006) we highlight three prevalent, and conflicting, themes: ‘same-sex marriage becomes legal under the Civil Partnership Act’; ‘couples will not get full legal status’ and ‘marriage is a heterosexual business’. We discuss these media representations and argue that the heteronormativity of the coverage provided little space for more radical constructions of same-sex relationship recognition.

A common rule for integration and suppression of luminance contrast across eyes, space, time, and pattern

Visual perception begins by dissecting the retinal image into millions of small patches for local analyses by local receptive fields. However, image structures extend well beyond these receptive fields and so further processes must be involved in sewing the image fragments back together to derive representations of higher order (more global) structures. To investigate the integration process, we also need to understand the opposite process of suppression. To investigate both processes together, we measured triplets of dipper functions for targets and pedestals involving interdigitated stimulus pairs (A, B). Previous work has shown that summation and suppression operate over the full contrast range for the domains of ocularity and space. Here, we extend that work to include orientation and time domains. Temporal stimuli were 15-Hz counter-phase sine-wave gratings, where A and B were the positive and negative phases of the oscillation, respectively. For orientation, we used orthogonally oriented contrast patches (A, B) whose sum was an isotropic difference of Gaussians. Results from all four domains could be understood within a common framework in which summation operates separately within the numerator and denominator of a contrast gain control equation. This simple arrangement of summation and counter-suppression achieves integration of various stimulus attributes without distorting the underlying contrast code.

Shared action spaces:a basis function framework for social re-calibration of sensorimotor representations supporting joint action

The article explores the possibilities of formalizing and explaining the mechanisms that support spatial and social perspective alignment sustained over the duration of a social interaction. The basic proposed principle is that in social contexts the mechanisms for sensorimotor transformations and multisensory integration (learn to) incorporate information relative to the other actor(s), similar to the "re-calibration" of visual receptive fields in response to repeated tool use. This process aligns or merges the co-actors' spatial representations and creates a "Shared Action Space" (SAS) supporting key computations of social interactions and joint actions; for example, the remapping between the coordinate systems and frames of reference of the co-actors, including perspective taking, the sensorimotor transformations required for lifting jointly an object, and the predictions of the sensory effects of such joint action. The social re-calibration is proposed to be based on common basis function maps (BFMs) and could constitute an optimal solution to sensorimotor transformation and multisensory integration in joint action or more in general social interaction contexts. However, certain situations such as discrepant postural and viewpoint alignment and associated differences in perspectives between the co-actors could constrain the process quite differently. We discuss how alignment is achieved in the first place, and how it is maintained over time, providing a taxonomy of various forms and mechanisms of space alignment and overlap based, for instance, on automaticity vs. control of the transformations between the two agents. Finally, we discuss the link between low-level mechanisms for the sharing of space and high-level mechanisms for the sharing of cognitive representations. © 2013 Pezzulo, Iodice, Ferraina and Kessler.

Visualising and controlling the flow in biomolecular systems at and between multiple scales:from atoms to hydrodynamics at different locations in time and space

A novel framework for modelling biomolecular systems at multiple scales in space and time simultaneously is described. The atomistic molecular dynamics representation is smoothly connected with a statistical continuum hydrodynamics description. The system behaves correctly at the limits of pure molecular dynamics (hydrodynamics) and at the intermediate regimes when the atoms move partly as atomistic particles, and at the same time follow the hydrodynamic flows. The corresponding contributions are controlled by a parameter, which is defined as an arbitrary function of space and time, thus, allowing an effective separation of the atomistic 'core' and continuum 'environment'. To fill the scale gap between the atomistic and the continuum representations our special purpose computer for molecular dynamics, MDGRAPE-4, as well as GPU-based computing were used for developing the framework. These hardware developments also include interactive molecular dynamics simulations that allow intervention of the modelling through force-feedback devices.

Normal ECG Recognition for Express-Diagnostics Based on Scale-Space Representation and Dynamic Matching

A novel approach of normal ECG recognition based on scale-space signal representation is proposed. The approach utilizes curvature scale-space signal representation used to match visual objects shapes previously and dynamic programming algorithm for matching CSS representations of ECG signals. Extraction and matching processes are fast and experimental results show that the approach is quite robust for preliminary normal ECG recognition.

Automatic ECG Analysis for Preliminary and Detailed Diagnostics Based on Scale-space Representation

A novel approach of automatic ECG analysis based on scale-scale signal representation is proposed. The approach uses curvature scale-space representation to locate main ECG waveform limits and peaks and may be used to correct results of other ECG analysis techniques or independently. Moreover dynamic matching of ECG CSS representations provides robust preliminary recognition of ECG abnormalities which has been proven by experimental results.

Representations of Inverse Functions by the Integral Transform with the Sign Kernel

Mathematics Subject Classification: Primary 30C40

One's own country and familiar places in the mind's eye:different topological representations for navigational and non-navigational contents

Visual mental imagery is a process that draws on different cognitive abilities and is affected by the contents of mental images. Several studies have demonstrated that different brain areas subtend the mental imagery of navigational and non-navigational contents. Here, we set out to determine whether there are distinct representations for navigational and geographical images. Specifically, we used a Spatial Compatibility Task (SCT) to assess the mental representation of a familiar navigational space (the campus), a familiar geographical space (the map of Italy) and familiar objects (the clock). Twenty-one participants judged whether the vertical or the horizontal arrangement of items was correct. We found that distinct representational strategies were preferred to solve different categories on the SCT, namely, the horizontal perspective for the campus and the vertical perspective for the clock and the map of Italy. Furthermore, we found significant effects due to individual differences in the vividness of mental images and in preferences for verbal versus visual strategies, which selectively affect the contents of mental images. Our results suggest that imagining a familiar navigational space is somewhat different from imagining a familiar geographical space. © 2014 Elsevier Ireland Ltd.

Continuum tensor network field states, path integral representations and spatial symmetries

A natural way to generalize tensor network variational classes to quantum field systems is via a continuous tensor contraction. This approach is first illustrated for the class of quantum field states known as continuous matrix-product states (cMPS). As a simple example of the path-integral representation we show that the state of a dynamically evolving quantum field admits a natural representation as a cMPS. A completeness argument is also provided that shows that all states in Fock space admit a cMPS representation when the number of variational parameters tends to infinity. Beyond this, we obtain a well-behaved field limit of projected entangled-pair states (PEPS) in two dimensions that provide an abstract class of quantum field states with natural symmetries. We demonstrate how symmetries of the physical field state are encoded within the dynamics of an auxiliary field system of one dimension less. In particular, the imposition of Euclidean symmetries on the physical system requires that the auxiliary system involved in the class' definition must be Lorentz-invariant. The physical field states automatically inherit entropy area laws from the PEPS class, and are fully described by the dissipative dynamics of a lower dimensional virtual field system. Our results lie at the intersection many-body physics, quantum field theory and quantum information theory, and facilitate future exchanges of ideas and insights between these disciplines.

Efficient Image and Video Representations for Retrieval

Image (Video) retrieval is an interesting problem of retrieving images (videos) similar to the query. Images (Videos) are represented in an input (feature) space and similar images (videos) are obtained by finding nearest neighbors in the input representation space. Numerous input representations both in real valued and binary space have been proposed for conducting faster retrieval. In this thesis, we present techniques that obtain improved input representations for retrieval in both supervised and unsupervised settings for images and videos. Supervised retrieval is a well known problem of retrieving same class images of the query. We address the practical aspects of achieving faster retrieval with binary codes as input representations for the supervised setting in the first part, where binary codes are used as addresses into hash tables. In practice, using binary codes as addresses does not guarantee fast retrieval, as similar images are not mapped to the same binary code (address). We address this problem by presenting an efficient supervised hashing (binary encoding) method that aims to explicitly map all the images of the same class ideally to a unique binary code. We refer to the binary codes of the images as `Semantic Binary Codes' and the unique code for all same class images as `Class Binary Code'. We also propose a new class­ based Hamming metric that dramatically reduces the retrieval times for larger databases, where only hamming distance is computed to the class binary codes. We also propose a Deep semantic binary code model, by replacing the output layer of a popular convolutional Neural Network (AlexNet) with the class binary codes and show that the hashing functions learned in this way outperforms the state­ of ­the art, and at the same time provide fast retrieval times. In the second part, we also address the problem of supervised retrieval by taking into account the relationship between classes. For a given query image, we want to retrieve images that preserve the relative order i.e. we want to retrieve all same class images first and then, the related classes images before different class images. We learn such relationship aware binary codes by minimizing the similarity between inner product of the binary codes and the similarity between the classes. We calculate the similarity between classes using output embedding vectors, which are vector representations of classes. Our method deviates from the other supervised binary encoding schemes as it is the first to use output embeddings for learning hashing functions. We also introduce new performance metrics that take into account the related class retrieval results and show significant gains over the state­ of­ the art. High Dimensional descriptors like Fisher Vectors or Vector of Locally Aggregated Descriptors have shown to improve the performance of many computer vision applications including retrieval. In the third part, we will discuss an unsupervised technique for compressing high dimensional vectors into high dimensional binary codes, to reduce storage complexity. In this approach, we deviate from adopting traditional hyperplane hashing functions and instead learn hyperspherical hashing functions. The proposed method overcomes the computational challenges of directly applying the spherical hashing algorithm that is intractable for compressing high dimensional vectors. A practical hierarchical model that utilizes divide and conquer techniques using the Random Select and Adjust (RSA) procedure to compress such high dimensional vectors is presented. We show that our proposed high dimensional binary codes outperform the binary codes obtained using traditional hyperplane methods for higher compression ratios. In the last part of the thesis, we propose a retrieval based solution to the Zero shot event classification problem - a setting where no training videos are available for the event. To do this, we learn a generic set of concept detectors and represent both videos and query events in the concept space. We then compute similarity between the query event and the video in the concept space and videos similar to the query event are classified as the videos belonging to the event. We show that we significantly boost the performance using concept features from other modalities.