Covariance matrices have attracted attention for machine learning applications due to their capacity capture interesting structure in the data. The main challenge is that one needs take into account particular geometry of Riemannian manifold symmetric positive definite (SPD) they belong to. In context deep networks, several architectures these recently been proposed. our article, we introduce a batch normalization (batchnorm) algorithm, which generalizes used Euclidean nets. This novel layer...

This work builds temporal deep learning architectures for the classification of time-frequency signal representations on a novel model simulated radar datasets. We show and compare success these models validate interest structures to gain confidence over time.

The problem of keyword spotting i.e. identifying keywords in a real-time audio stream is mainly solved by applying neural network over successive sliding windows. Due to the difficulty task, baseline models are usually large, resulting high computational cost and energy consumption level. We propose new method called SANAS (Stochastic Adaptive Neural Architecture Search) which able adapt architecture on-the-fly at inference time such that small architectures will be used when easy process...

Micro-Doppler analysis commonly makes use of the log-scaled, real-valued spectrogram, and recent work involving deep learning architectures for classification are no exception. Some works in neighboring fields research directly exploit raw temporal signal, but do not handle complex numbers, which inherent to radar IQ signals. In this paper, we propose a complex-valued, fully neural network simultaneously exploits signal spectrogram by introducing Fourier-like layer suitable architectures. We...

Mesoscale oceanic eddies have a visible signature on sea surface temperature (SST) satellite images, portraying diverse patterns of coherent vortices, gradients, and swirling filaments. However, learning the regularities such signatures defines challenging pattern recognition task, due to their complex structure but also cloud coverage which can corrupt large fraction image. We introduce novel deep approach classify eddy signatures, even if they are corrupted by strong coverage. A dataset...

Until now, mesoscale oceanic eddies have been automatically detected through physical methods on satellite altimetry. Nevertheless, they often a visible signature Sea Surface Temperature (SST) images, which not yet sufficiently exploited. We introduce novel method that employs Deep Learning to detect eddy signatures such input. provide the first available dataset for this task, retaining SST images altimetric-based region proposal. train CNN-based classifier succeeds in accurately detecting...

Gaussian mixture models are a widespread tool for modeling various and complex probability density functions. They can be estimated using Expectation- Maximization or Kernel Density Estimation. leads to compact but may expensive compute whereas Estimation yields large which cheap build. In this paper we present new methods get high-quality that both fast compute. This is accomplished with clustering centroids computation. The quality of the resulting mixtures evaluated in terms...

Classification of radar observations with machine learning tools is primary importance for the identification non-cooperative targets such as drones. These are made complex-valued time series which possess a strong underlying structure. signals can be processed through time-frequency analysis, their self-correlation (or covariance) matrices or directly raw signal. All representations linked but distinct and it known that input representation critical success any method. In this article, we...

In this work, we build dedicated learning models for micro-Doppler radar time series classification. We develop both deep temporal architectures based on time-frequency representations, and also directly study the signal's underlying statistical Gaussian process using Information Geometry Riemannian manifolds by developing improving symmetric positive definite (SPD) neural networks. propose aggregation of all proposed in a single, highly performing classification pipeline.

Bhattacharrya distance (BD) is a widely used in statistics to compare probability density functions (PDFs). It has shown strong statistical properties (in terms of Bayes error) and it relates Fisher information. also practical advantages, since strongly on measuring the overlap supports PDFs. Unfortunately, even with common parametric models PDFs, few closed-form formulas are known. Moreover, BD centroid estimation was limited univariate gaussian PDFs literature no convergence guarantees...

We present a novel reranking framework for Content Based Image Retrieval (CBIR) systems based on contextual dissimilarity measures. Our work revisit and extend the method of Perronnin et al. (Perronnin al., 2009) which introduces way to build contexts used in turn design measures reranking. Instead using truncated rank lists from CBIR engine as contexts, we rather use clustering algorithm group similar images list. introduce representational Bregman divergences further generalize k-means by...

Ribonucleic acid (RNA) molecules play important roles in a variety of biological processes. To properly function, RNA usually have to fold specific structures, and therefore understanding structure is vital comprehending how functions. One approach predicting biomolecular use knowledge-based potentials built from experimentally determined structures. These types been shown be effective for both protein but their utility limited by significantly rugged nature. This ruggedness (and hence the...

The Kullback-Leibler divergence is a widespread dissimilarity measure between probability density functions, based on the Shannon entropy. Unfortunately, there no analytic formula available to compute this mixture models, imposing use of costly approximation algorithms. In order reduce computational burden when lot evaluations are needed, we introduce sub-class models where component parameters shared set mixtures and only degree-of-freedom vector weights each mixture. This sharing allows...

In its raw form, micro-Doppler radar data takes the form of a complex time-series, which can be seen as multiple realizations Gaussian process. As such, covariance matrix constitutes viable and synthetic representation such data. this paper, we introduce neural network on Hermitian Positive Definite (HPD) matrices, that is complex-valued Symmetric (SPD) or matrices. We validate new architecture data, comparing against previous similar methods.

The scope of the well-known k-means algorithm has been broadly extended with some recent results: first, k means++ initialization method gives approximation guarantees; second, Bregman generalizes classical to large family divergences. seeding framework combines guarantees We present here an extension using α-divergences. With for representational divergences, we show that α-divergence based can be designed. preliminary experiments clustering and image segmentation applications. Since...

Modeling data is often a critical step in many challenging applications computer vision, bioinformatics or machine learning. Gaussian Mixture Models are popular choice applications. Although these mixtures powerful enough to approximate complex distributions, they may not be the best for some Usual software libraries limited particular kind of distribution, which makes difficult change distribution and so choose one. In this paper we focus on class exponential families (which contains lot...

Surface currents provided, in real time, by operational ocean models often differ from each other but also satellite altimetry observations, especially terms of mesoscale dynamics. Eddies, which play a dominant role on circulation at the regional scale, have signature both maps and imagery, such as sea surface temperature. Combining these independent signatures allows for highly reliable detection reference eddies. To this end, we build convolutional neural network capable detecting contours...