Everyday an enormous amount of information is stored, processed and transmitted digitally around the world. Neurally-inspired compression models have been rapidly developed and researched as a solution to image processing tasks and channel error correction control. This dissertation presents a deep neural network (DNN) for gray high-resolution image compression and a fault-tolerant transmission system with channel error-correction capabilities. A feed-forward DNN implemented with the Levenberg-Marguardt learning algorithm is proposed and implemented for image compression. I demonstrate experimentally that the DNN not only provides better quality reconstructed images but also requires less computational capacity as compared to DCT Zonal coding, DCT Threshold coding, Set Partitioning in Hierarchical Trees (SPIHT) and Gaussian Pyramid. An artificial neural network (ANN) with improved channel error-correction rate is also proposed. The experimental results indicate that the implemented artificial neural network provides a superior error-correction ability by transmitting binary images over the noisy channel using Hamming and Repeat-Accumulate coding. Meanwhile, the network’s storage requirement is 64 times less than the Hamming coding and 62 times less than the Repeat-Accumulate coding. Thumbnail images contain higher frequencies and much less redundancy, which makes them more difficult to compress compared to high-resolution images. Bottleneck autoencoders have been actively researched as a solution to image compression tasks. However, I observed that thumbnail images compressed at a 2:1 ratio through bottleneck autoencoders often exhibit subjectively low visual quality. In this dissertation, I compared bottleneck autoencoders with two sparse coding approaches. Either 50\% of the pixels are randomly removed or every other pixel is removed, each achieving a 2:1 compression ratio. In the subsequent decompression step, a sparse inference algorithm is used to in-paint the missing the pixel values. Compared to bottleneck autoencoders, I observed that sparse coding with a random dropout mask yields decompressed images that are superior based on subjective human perception yet inferior according to pixel-wise metrics of reconstruction quality, such as PSNR and SSIM. With a regular checkerboard mask, decompressed images were superior as assessed by both subjective and pixel-wise measures. I hypothesized that alternative feature-based measures of reconstruction quality would better support my subjective observations. To test this hypothesis, I fed thumbnail images processed using either bottleneck autoencoder or sparse coding using either checkerboard or random masks into a Deep Convolutional Neural Network (DCNN) classifier. Consistent, with my subjective observations, I discovered that sparse coding with checkerboard and random masks support on average 2.7\% and 1.6\% higher classification accuracy and 18.06\% and 3.74\% lower feature perceptual loss compared to bottleneck autoencoders, implying that sparse coding preserves more feature-based information. The optic nerve transmits visual information to the brain as trains of discrete events, a low-power, low-bandwidth communication channel also exploited by silicon retina cameras. Extracting high-fidelity visual input from retinal event trains is thus a key challenge for both computational neuroscience and neuromorphic engineering. % Here, we investigate whether sparse coding can enable the reconstruction of high-fidelity images and video from retinal event trains. Our approach is analogous to compressive sensing, in which only a random subset of pixels are transmitted and the missing information is estimated via inference. We employed a variant of the Locally Competitive Algorithm to infer sparse representations from retinal event trains, using a dictionary of convolutional features optimized via stochastic gradient descent and trained in an unsupervised manner using a local Hebbian learning rule with momentum. Static images, drawn from the CIFAR10 dataset, were passed to the input layer of an anatomically realistic retinal model and encoded as arrays of output spike trains arising from separate layers of integrate-and-fire neurons representing ON and OFF retinal ganglion cells. The spikes from each model ganglion cell were summed over a 32 msec time window, yielding a noisy rate-coded image. Analogous to how the primary visual cortex is postulated to infer features from noisy spike trains in the optic nerve, we inferred a higher-fidelity sparse reconstruction from the noisy rate-coded image using a convolutional dictionary trained on the original CIFAR10 database. Using a similar approach, we analyzed the asynchronous event trains from a silicon retina camera produced by self-motion through a laboratory environment. By training a dictionary of convolutional spatiotemporal features for simultaneously reconstructing differences of video frames (recorded at 22HZ and 5.56Hz) as well as discrete events generated by the silicon retina (binned at 484Hz and 278Hz), we were able to estimate high frame rate video from a low-power, low-bandwidth silicon retina camera.

abstract: Alzheimer’s disease (AD), is a chronic neurodegenerative disease that usually starts slowly and gets worse over time. It is the cause of 60% to 70% of cases of dementia. There is growing interest in identifying brain image biomarkers that help evaluate AD risk pre-symptomatically. High-dimensional non-linear pattern classification methods have been applied to structural magnetic resonance images (MRI’s) and used to discriminate between clinical groups in Alzheimers progression. Using Fluorodeoxyglucose (FDG) positron emission tomography (PET) as the pre- ferred imaging modality, this thesis develops two independent machine learning based patch analysis methods and uses them to perform six binary classification experiments across different (AD) diagnostic categories. Specifically, features were extracted and learned using dimensionality reduction and dictionary learning & sparse coding by taking overlapping patches in and around the cerebral cortex and using them as fea- tures. Using AdaBoost as the preferred choice of classifier both methods try to utilize 18F-FDG PET as a biological marker in the early diagnosis of Alzheimer’s . Addi- tional we investigate the involvement of rich demographic features (ApoeE3, ApoeE4 and Functional Activities Questionnaires (FAQ)) in classification. The experimental results on Alzheimer’s Disease Neuroimaging initiative (ADNI) dataset demonstrate the effectiveness of both the proposed systems. The use of 18F-FDG PET may offer a new sensitive biomarker and enrich the brain imaging analysis toolset for studying the diagnosis and prognosis of AD. / Dissertation/Thesis / Thesis Defense Presentation / Masters Thesis Computer Science 2017

Les ondes acoustiques subissent peu de dispersion dans le milieu marin, comparé au milieu aérien. Certaines espèces de cétacés communiquent ainsi à grande distance, d'autres utilisent leurs émissions sonores pour s'orienter. La bioacoustique consiste à étudier ces espèces à partir de l'analyse de leurs sons, c'est-à-dire à les détecter, classer, localiser. Cela peut se faire via un réseau d'hydrophones au déploiement fastidieux. Afin de contribuer au passage à l'échelle de la bioacoustique, cette thèse propose des modèles originaux mono-hydrophone pour l'analyse de ces signaux stationnaires ou transitoires. Premièrement, nous dérivons un nouveau modèle d'estimation de la distance entre une source impulsive (ex. biosonar) et un hydrophone. Notre modèle théorique, l'Intra Spectral ATténuation(ISAT), dérive des lois acoustiques de déformation spectrale du signal transitoire induite par l'atténuation durant sa propagation. Ce modèle relie les rapports énergétiques des bandes de fréquences pondérés par un modèle de perte par atténuation fréquentielle (Thorp ou Leroy) à la distance de propagation. Nous approximons aussi ISAT par un modèle neuromimétique. Ces deux modèles sont validés sur le sonar du cachalot (Physeter macrocephalus) enregistré avec notre bouée acoustique autonome BOMBYX et notre système d'acquisition DECAV en collaboration avec le Parc National de Port-Cros et le sanctuaire Pelagos pour la protection des mammifères marins en Méditerranée. Les mesures d'erreur (RMSE) d'environ 500 mètres sur nos références du centre d'essai OTAN aux Bahamas présentent un intérêt opérationnel. Deuxièmement, nous proposons une analyse originale de l'évolution des voisements de cétacé par codage parcimonieux. Notre encodage des cepstres par apprentissage non supervisé d'un dictionnaire met en évidence l'évolution temporelle des bigrammes des chants que les baleines à bosse mâles émettent durant la période de reproduction. Nous validons ce modèle sur nos enregistrements du canal de Sainte-Marie à Madagascar entre 2008 et 2014, via notre réseau d'hydrophones BAOBAB qui constitue une première dans l'Océan Indien. Nos modèles s'inscrivent dans le projet Scaled Bioacoustics (SABIOD, MI CNRS) et ouvrent de nouvelles perspectives pour les passages à l'échelle temporelle et spatiale de la bioacoustique. / Acoustic waves show low dispersion due to the underwater propagation, compared to the propagation in the air. Some species of cetaceans communicate at long distance, others use their sound production for orientation. The goal of the scientic area called bioacoustics is to study animal species based on the analysis of their emitted sound. Their sounds can be used to detect, to classify and to locate the cetaceans. Recordings can be done with an passive acoustic array of multiple hydrophones, but this method is expensive and difficult to deploy. Thus, in order to scale this approach, we propose in this Phd thesis several original single hydrophone models to analyze these stationary or transient signals.Firstly, we provide a new theoretical model to estimate the distance between the impulsive source (ex. biosonar of the cetacean) and the hydrophone. Our model, the Intra Spectral ATtenuation (ISAT), is based on the spectral signal alteration due to the underwater acoustic propagation, especially the differences in different frequency bands. We also approximated ISAT by an artificial neural network. Both models are validated on clicks emitted by sperm whales (Physeter macrocephalus) recorded by our sonobuoy BOMBYX and our data-acquisition system DECAV developed incollaboration with the National Park of Port-Cros (France) and the Pelagos sanctuary for the protection of marine mammals in the Mediterranean sea (France). The error (RMSE) measures on the recordings of the NATO test center in the Bahamas are about500 meters, promising further real applications. Secondly, we worked on the variations of the cetacean vocalizations using the sparse coding method. The encoding of thecepstrums by unsupervised learning of a dictionary shows bigrammic time changes of the songs of humpback whales (Megaptera novaeangliae). We validate this model on signals recorded in the Ste Marie Channel (Madagascar) between 2008 and 2014, through our network of hydrophones BAOBAB which is the first passive acoustic array deployed in the Indian Ocean.Our models are part of the Saled Bioacoustics project (SABIOD, MI CNRS) and open perspectives for temporal and spatial scaling of bioacoustics.

Graphs model relationships across real-world entities in web graphs, social network graphs, and road network graphs. Graph algorithms analyze and transform a graph to discover graph properties or to apply a computation. For instance, a pagerank algorithm computes a rank for each page in a webgraph, and a community detection algorithm discovers likely communities in a social network, while a shortest path algorithm computes the quickest way to reach a place from another, in a road network. In Domains such as social information systems, the number of edges can be in billions or trillions. Such large graphs are processed on distributed computer systems or clusters. Graph algorithms can be executed on multi-core CPUs, GPUs with thousands of cores, multi-GPU devices, and CPU+GPU clusters, depending on the size of the graph object. While programming such algorithms on heterogeneous targets, a programmer is required to deal with parallelism and and also manage explicit data communication between distributed devices. This implies that a programmer is required to learn CUDA, OpenMP, MPI, etc., and also the details of the hardware architecture. Such codes are error prone and di cult to debug. A Domain Speci c Language (DSL) which hides all the hardware details and lets the programmer concentrate only the algorithmic logic will be very useful. With this as the research goal, Falcon, graph DSL and its compiler have been developed. Falcon programs are explicitly parallel and Falcon hides all the hardware details from the programmer. Large graphs that do not t into the memory of a single device are automatically partitioned by the Falcon compiler. Another feature of Falcon is that it supports mutation of graph objects and thus enables programming dynamic graph algorithms. The Falcon compiler converts a single DSL code to heterogeneous targets such as multi-core CPUs, GPUs, multi-GPU devices, and CPU+GPU clusters. Compiled codes of Falcon match or outperform state-of-the-art graph frameworks for di erent target platforms and benchmarks.

Cette thèse explore l'analyse de l'activité électrique du cerveau. Un défi important de ces signaux est leur grande variabilité à travers différents essais et/ou différents sujets. Nous proposons une nouvelle méthode appelée "adaptive waveform learning" (AWL). Cette méthode est suffisamment générale pour permettre la prise en compte de la variabilité empiriquement rencontrée dans les signaux neuroélectriques, mais peut être spécialisée afin de prévenir l'overfitting du bruit. La première partie de ce travail donne une introduction sur l'électrophysiologie du cerveau, présente les modalités d'enregistrement fréquemment utilisées et décrit l'état de l'art du traitement de signal neuroélectrique. La principale contribution de cette thèse consiste en 3 chapitres introduisant et évaluant la méthode AWL. Nous proposons d'abord un modèle de décomposition de signal général qui inclut explicitement différentes formes de variabilité entre les composantes de signal. Ce modèle est ensuite spécialisé pour deux applications concrètes: le traitement d'une série d'essais expérimentaux segmentés et l'apprentissage de structures répétées dans un seul signal. Deux algorithmes sont développés pour résoudre ces problèmes de décomposition. Leur implémentation efficace basée sur des techniques de minimisation alternée et de codage parcimonieux permet le traitement de grands jeux de données.Les algorithmes proposés sont évalués sur des données synthétiques et réelles contenant des pointes épileptiformes. Leurs performances sont comparées à celles de la PCA, l'ICA, et du template-matching pour la détection de pointe. / This thesis investigates the analysis of brain electrical activity. An important challenge is the presence of large variability in neuroelectrical recordings, both across different subjects and within a single subject, for example, across experimental trials. We propose a new method called adaptive waveform learning (AWL). It is general enough to include all types of relevant variability empirically found in neuroelectric recordings, but can be specialized for different concrete settings to prevent from overfitting irrelevant structures in the data. The first part of this work gives an introduction into the electrophysiology of the brain, presents frequently used recording modalities, and describes state-of-the-art methods for neuroelectrical signal processing. The main contribution of this thesis consists in three chapters introducing and evaluating the AWL method. We first provide a general signal decomposition model that explicitly includes different forms of variability across signal components. This model is then specialized for two concrete applications: processing a set of segmented experimental trials and learning repeating structures across a single recorded signal. Two algorithms are developed to solve these models. Their efficient implementation based on alternate minimization and sparse coding techniques allows the processing of large datasets. The proposed algorithms are evaluated on both synthetic data and real data containing epileptiform spikes. Their performances are compared to those of PCA, ICA, and template matching for spike detection.

Les représentations convolutives extraient des motifs récurrents qui aident à comprendre la structure locale dans un jeu de signaux. Elles sont adaptées pour l’analyse des signaux physiologiques, qui nécessite des visualisations mettant en avant les informations pertinentes. Ces représentations sont aussi liées aux modèles d’apprentissage profond. Dans ce manuscrit, nous décrivons des avancées algorithmiques et théoriques autour de ces modèles. Nous montrons d’abord que l’Analyse du Spectre Singulier permet de calculer efficacement une représentation convolutive. Cette représentation est dense et nous décrivons une procédure automatisée pour la rendre plus interprétable. Nous proposons ensuite un algorithme asynchrone, pour accélérer le codage parcimonieux convolutif. Notre algorithme présente une accélération super-linéaire. Dans une seconde partie, nous analysons les liens entre représentations et réseaux de neurones. Nous proposons une étape d’apprentissage supplémentaire, appelée post-entraînement, qui permet d’améliorer les performances du réseau entraîné, en s’assurant que la dernière couche soit optimale. Puis nous étudions les mécanismes qui rendent possible l’accélération du codage parcimonieux avec des réseaux de neurones. Nous montrons que cela est lié à une factorisation de la matrice de Gram du dictionnaire. Finalement, nous illustrons l’intérêt de l’utilisation des représentations convolutives pour les signaux physiologiques. L’apprentissage de dictionnaire convolutif est utilisé pour résumer des signaux de marche et le mouvement du regard est soustrait de signaux oculométriques avec l’Analyse du Spectre Singulier. / Convolutional representations extract recurrent patterns which lead to the discovery of local structures in a set of signals. They are well suited to analyze physiological signals which requires interpretable representations in order to understand the relevant information. Moreover, these representations can be linked to deep learning models, as a way to bring interpretability intheir internal representations. In this disserta tion, we describe recent advances on both computational and theoretical aspects of these models.First, we show that the Singular Spectrum Analysis can be used to compute convolutional representations. This representation is dense and we describe an automatized procedure to improve its interpretability. Also, we propose an asynchronous algorithm, called DICOD, based on greedy coordinate descent, to solve convolutional sparse coding for long signals. Our algorithm has super-linear acceleration.In a second part, we focus on the link between representations and neural networks. An extra training step for deep learning, called post-training, is introduced to boost the performances of the trained network by making sure the last layer is optimal. Then, we study the mechanisms which allow to accelerate sparse coding algorithms with neural networks. We show that it is linked to afactorization of the Gram matrix of the dictionary.Finally, we illustrate the relevance of convolutional representations for physiological signals. Convolutional dictionary learning is used to summarize human walk signals and Singular Spectrum Analysis is used to remove the gaze movement in young infant’s oculometric recordings.

Dans les séries temporelles neurophysiologiques, on observe de fortes oscillations neuronales, et les outils d'analyse sont donc naturellement centrés sur le filtrage à bande étroite.Puisque cette approche est trop réductrice, nous proposons de nouvelles méthodes pour représenter ces signaux.Nous centrons tout d'abord notre étude sur le couplage phase-amplitude (PAC), dans lequel une bande haute fréquence est modulée en amplitude par la phase d'une oscillation neuronale plus lente.Nous proposons de capturer ce couplage dans un modèle probabiliste appelé modèle autoregressif piloté (DAR). Cette modélisation permet une sélection de modèle efficace grâce à la mesure de vraisemblance, ce qui constitue un apport majeur à l'estimation du PAC.%Nous présentons différentes paramétrisations des modèles DAR et leurs algorithmes d'inférence rapides, et discutons de leur stabilité.Puis nous montrons comment utiliser les modèles DAR pour l'analyse du PAC, et démontrons l'avantage de l'approche par modélisation avec trois jeux de donnée.Puis nous explorons plusieurs extensions à ces modèles, pour estimer le signal pilote à partir des données, le PAC sur des signaux multivariés, ou encore des champs réceptifs spectro-temporels.Enfin, nous proposons aussi d'adapter les modèles de codage parcimonieux convolutionnels pour les séries temporelles neurophysiologiques, en les étendant à des distributions à queues lourdes et à des décompositions multivariées. Nous développons des algorithmes d'inférence efficaces pour chaque formulations, et montrons que l'on obtient de riches représentations de façon non-supervisée. / In neurophysiological time series, strong neural oscillations are observed in the mammalian brain, and the natural processing tools are thus centered on narrow-band linear filtering.As this approach is too reductive, we propose new methods to represent these signals.We first focus on the study of phase-amplitude coupling (PAC), which consists in an amplitude modulation of a high frequency band, time-locked with a specific phase of a slow neural oscillation.We propose to use driven autoregressive models (DAR), to capture PAC in a probabilistic model. Giving a proper model to the signal enables model selection by using the likelihood of the model, which constitutes a major improvement in PAC estimation.%We first present different parametrization of DAR models, with fast inference algorithms and stability discussions.Then, we present how to use DAR models for PAC analysis, demonstrating the advantage of the model-based approach on three empirical datasets.Then, we explore different extensions to DAR models, estimating the driving signal from the data, PAC in multivariate signals, or spectro-temporal receptive fields.Finally, we also propose to adapt convolutional sparse coding (CSC) models for neurophysiological time-series, extending them to heavy-tail noise distribution and multivariate decompositions. We develop efficient inference algorithms for each formulation, and show that we obtain rich unsupervised signal representations.

No description available.

No description available.

Die Aktionspotential (AP) -Aktivität einzelner kortikaler Neuronen kann messbare sensorische Effekte hervorrufen. Es ist jedoch nicht bekannt, wie AP-Sequenzen Parameter und spezifische neuronale Subtypen die hervorgerufenen Sinnesempfindungen beeinflussen. Hier haben wir einen ‘Reverse-Physiology‘ Ansatz angewendet, um die Beziehung zwischen der Aktivität einzelner Neuronen und der Empfindung zu untersuchen. Zunächst wird der Prozess der Nanostimulation, eine von der juxtazellulären Markierungstechnik abgeleiteten Einzelzell-Stimulationsmethode, detailliert beschrieben. Nanostimulation ist einfach anzuwenden und kann auf eine Vielzahl von identifizierbaren Neuronen in narkotisierten und wachen Tieren angewandt werden. Wir beschreiben die Aufnahmetechnik und die elektrische Konfiguration für Nanostimulation. Während eine exakte zeitliche Bestimmung der AP nicht erreicht wurde, konnten Frequenz und Anzahl der AP parametrisch kontrolliert werden. Wir zeigen, dass Nanostimulation auch angewendet werden kann, um sensorische Reaktionen in identifizierbaren Neuronen selektiv zu inhibieren. Als nächstes haben wir untersucht wie sich die Frequenz und Anzahl der AP sowie die Regelmäßigkeit der Pulsfolge auf die Detektion von Einzelzell-Stimulationen im somatosensorischen Kortex von Ratten auswirken. Für mutmaßlichen erregende regular-spiking Neuronen erhöhte sich die Nachweisbarkeit mit abnehmender Frequenz und Anzahl der AP. Die Stimulation einzelner, mutmaßlichen inhibitorischer und schnell feuernder Neuronen führte zu wesentlich stärkeren sensorischen Effekten, die unabhängig von Frequenz und Anzahl der AP waren. Außerdem fanden wir heraus, dass Unregelmäßigkeiten der Pulsfolge die sensorischen Effekte von putativ erregenden Neuronen stark erhöhten. Diese Unregelmäßigkeiten wurden in durchschnittlich 8% der Durchgänge festgestellt. Unsere Daten deuten darauf hin, dass das es auf Verhaltnisebene eine große Sensivität für kortikale AP und deren zeitlichen Abfolge gibt. / The action potential (AP) activity of single cortical neurons can evoke measurable sensory effects, but it is not known how spiking parameters and specific neuronal subtypes affect the evoked sensations. Here we applied a reverse physiology approach to investigate the relationship between single neuron activity and sensation. First, we provide a detailed description of the procedures involved in nanostimulation, a single-cell stimulation method derived from the juxtacellular labeling technique. Nanostimulation is easy to apply and can be directed to a wide variety of identifiable neurons in anesthetized and awake animals. We describe the recording approach and the parameters of the electric configuration underlying nanostimulation. While exact AP timing has not been achieved, AP frequency and AP number can be parametrically controlled. We demonstrate that nanostimulation can also be used to selectively inhibit sensory responses in identifiable neurons. Next, we examined the effects of AP frequency, AP number and spike train regularity on the detectability of single-cell stimulation in rat somatosensory cortex. For putative excitatory, regular spiking neurons detectability increased with decreasing AP frequencies and decreasing AP numbers. Stimulation of single putative inhibitory, fast spiking neurons led to much larger sensory effects that were not dependent on AP frequency and AP number. In addition, we found that spike train irregularity greatly increased the sensory effects of putative excitatory neurons, with irregular spike trains being detected in on average 8% of trials. Our data suggest that the behaving animal is extremely sensitive to cortical APs and their temporal patterning.