Search

Machine learning techniques typically used in dementia assessment are not able to learn multiple tasks jointly and deal with time-dependent heterogeneous data containing missing values. In this paper, we reformulate SSHIBA, a recently introduced Bayesian multi-view latent variable model, for jointly learning diagnosis, ventricle volume, and ADAS score in dementia on longitudinal data with missing values.We propose a novel Bayesian Variational inference framework capable of simultaneously imputing missing values and combining information from several views. This way, we can combine different data views from different time-points in a common latent space and learn the relationships between each time-point, using the semi-supervised formulation to fully exploit the temporal structure of the data and handle missing values. In turn, the model can combine all the available information to simultaneously model and predict multiple output variables.We applied the proposed model to jointly predict diagnosis, ventricle volume, and ADAS score in dementia. The comparison of imputation strategies demonstrated the superior performance of the semi-supervised formulation of the model, improving the best baseline methods. Moreover, the performance in simultaneous prediction of diagnosis, ventricle volume, and ADAS score led to an improved prediction performance over the best baseline method.The results demonstrate that the proposed SSHIBA framework can learn an excellent imputation of the missing values and outperforming the baselines while simultaneously predicting three different tasks.

We present a novel deep generative model based on non i.i.d. variational autoencoders that captures global dependencies among observations in a fully unsupervised fashion. In contrast to the recent semi-supervised alternatives for global modeling in deep generative models, our approach combines a mixture model in the local or data-dependent space and a global Gaussian latent variable, which lead us to obtain three particular insights. First, the induced latent global space captures interpretable disentangled representations with no user-defined regularization in the evidence lower bound (as in beta-VAE and its generalizations). Second, we show that the model performs domain alignment to find correlations and interpolate between different databases. Finally, we study the ability of the global space to discriminate between groups of observations with non-trivial underlying structures, such as face images with shared attributes or defined sequences of digits images. This work has been partly supported by Spanish government (AEI/MCI) under grants PID2021-123182OB-I00, PID2021-125159NB-I00 and RTI2018-099655-B-100, by Comunidad de Madrid under grant IND2022/TIC-23550, by the European Union (FEDER) and the European Research Council (ERC) through the European Union's Horizon 2020 research and innovation program under Grant 714161, and by Comunidad de Madrid and FEDER through IntCARE-CM. The work of Ignacio Peis has been also supported by by Spanish government (MIU) under grant FPU18/00516.

Matrix-Assisted Laser Desorption Ionization Time-Of-Flight (MALDI-TOF) Mass Spectrometry (MS) is a reference method for microbial identification and it can be used to predict Antibiotic Resistance (AR) when combined with artificial intelligence methods. However, current solutions need time-costly preprocessing steps, are difficult to reproduce due to hyperparameter tuning, are hardly interpretable, and do not pay attention to epidemiological differences inherent to data coming from different centres, which can be critical.We propose using a multi-view heterogeneous Bayesian model (KSSHIBA) for the prediction of AR using MALDI-TOF MS data together with their epidemiological differences. KSSHIBA is the first model that removes the ad-hoc preprocessing steps that work with raw MALDI-TOF data. In addition, due to its Bayesian probabilistic nature, it does not require hyperparameter tuning, provides interpretable results, and allows exploiting local epidemiological differences between data sources. To test the proposal, we used data from 402Klebsiella pneumoniaeisolates coming from two different domains and 20 different hospitals located in Spain and Portugal. KSSHIBA outperforms current state-of-the-art approaches in antibiotic susceptibility prediction, obtaining a 0.78 AUC score in Wild Type classification and a 0.90 AUC score in Extended-Spectrum Beta-Lactamases (ESBL)+Carbapenemases (CP)-producers. The proposal consistently removes the need for ad-hoc preprocessing by working with raw MALDI-TOF data, which, in turn, reduces the time needed to obtain the results of the resistance mechanism in microbiological laboratories. The proposed model implementation as well as both data domains are publicly available.

Time series forecasting is an important problem across many domains, playing a crucial role in multiple real-world applications. In this paper, we propose a forecasting architecture that combines deep autoregressive models with a Spectral Attention (SA) module, which merges global and local frequency domain information in the model's embedded space. By characterizing in the spectral domain the embedding of the time series as occurrences of a random process, our method can identify global trends and seasonality patterns. Two spectral attention models, global and local to the time series, integrate this information within the forecast and perform spectral filtering to remove time series's noise. The proposed architecture has a number of useful properties: it can be effectively incorporated into well-know forecast architectures, requiring a low number of parameters and producing interpretable results that improve forecasting accuracy. We test the Spectral Attention Autoregressive Model (SAAM) on several well-know forecast datasets, consistently demonstrating that our model compares favorably to state-of-the-art approaches., Errors in Eq. 2 and Eq. 3 corrected

In this paper, we analyze the tradeoff between coding rate and asymptotic performance of a class of generalized low-density parity-check (GLDPC) codes constructed by including a certain fraction of generalized constraint (GC) nodes in the graph. The rate of the GLDPC ensemble is bounded using classical results on linear block codes, namely Hamming bound and Varshamov bound. We also study the impact of the decoding method used at GC nodes. To incorporate both bounded-distance (BD) and Maximum Likelihood (ML) decoding at GC nodes into our analysis without resorting on multi-edge type of degree distributions (DDs), we propose the probabilistic peeling decoding (P-PD) algorithm, which models the decoding step at every GC node as an instance of a Bernoulli random variable with a successful decoding probability that depends on both the GC block code as well as its decoding algorithm. The P-PD asymptotic performance over the BEC can be efficiently predicted using standard techniques for LDPC codes such as density evolution (DE) or the differential equation method. Furthermore, for a class of GLDPC ensembles, we demonstrate that the simulated P-PD performance accurately predicts the actual performance of the GLPDC code under ML decoding at GC nodes. We illustrate our analysis for GLDPC code ensembles with regular and irregular DDs. In all cases, we show that a large fraction of GC nodes is required to reduce the original gap to capacity, but the optimal fraction is strictly smaller than one. We then consider techniques to further reduce the gap to capacity by means of random puncturing, and the inclusion of a certain fraction of generalized variable nodes in the graph., Submitted to IEEE Transactions on Information Theory, August 2017

In this work, a compact affordable and portable spectral imaging system is presented. The system is intended to be employed in general applications, such as material classification or determination of the concentration of chemical species together with colorimetric sensors. The imaging device is reduced to a small digital color detector with an active area of 3 × 2mm2. This device provides a quantification of the incident emission in the form of four digital words corresponding to its averaged components blue, green, red, and near infrared. In this way, the size of the image is reduced to one pixel. The wavelength selection is carried out by means of a LED array disposed surrounding the color detector. The LEDs are selected to cover the wavelength range from 360 to 890 nm. A sequential measurement protocol is followed, and the generated data is transmitted to an external portable device via a Bluetooth link where a classification protocol is implemented in a custom-developed Android™ application. The presented system has been applied in three different scenarios involving material classification, meat freshness monitoring, and chemical analysis. The analysis of the data using principal components shows that it is possible to find a set of wavelengths where the classification of the samples is optimal., This work was supported by project CTQ2016-78754-C2-1-R from the Spanish Ministry of Economy and Competitiveness.

In this chapter, we introduce Gaussian processes for machine learning and their application to designing digital communication receivers. Gaussian processes for machine learning are Bayesian nonlinear tools for solving regression and classification problems. Gaussian processes for regression (GPR) were introduced in the mid-nineties to solve nonparametric estimation problems from a Bayesian perspective. They place a Gaussian process (GP) prior over the possible regressors and use the available data to obtain a posterior regressor, which it is able to explain the observations without overfitting. The covariance matrix of the GP prior describes the different solutions that can be achieved, e.g. linear, polynomial, or universal regressors. The solution of GPR is analytical given its covariance function and, besides providing point estimates, it also assigns confidence intervals for the predictions. Furthermore, the covariance function can be optimized by maximum likelihood to better represent the data, which adds additional flexibility to our regression approximation. GPR can be generalized to solve classification problems, namely Gaussian processes for classification (GPC). GPC extends the idea of GPR for a classification likelihood model. For this likelihood, the GPC posterior is no longer analytically tractable and we need to approximate it. Expectation Propagation (EP), which matches the mean and covariance of the GP posterior to a Gaussian distribution, is the most widely used approximation. Unlike most state-of-the-art classifiers, GPC does not return point-wise decisions, but it provides an accurate posterior probability for each classification decision. This is a major advantage to be exploited by subsequent applications for reducing the base error produced by our nonlinear classifiers. Nonlinear regression and classification techniques have been widely used for designing digital communication receivers for nonlinear channels or whenever there is little information about the channel model or for nonlinear model. These nonlinear tools must use short training sequences to learn the channel and to adapt to a wide range of scenarios, from linear minimum phase to nonlinear and non-minimum phase and from single to multi-user 11

Medical data sets are usually corrupted by noise and missing data. These missing patterns are commonly assumed to be completely random, but in medical scenarios, the reality is that these patterns occur in bursts due to sensors that are off for some time or data collected in a misaligned uneven fashion, among other causes. This paper proposes to model medical data records with heterogeneous data types and bursty missing data using sequential variational autoencoders (VAEs). In particular, we propose a new methodology, the Shi-VAE, which extends the capabilities of VAEs to sequential streams of data with missing observations. We compare our model against state-of-the-art solutions in an intensive care unit database (ICU) and a dataset of passive human monitoring. Furthermore, we find that standard error metrics such as RMSE are not conclusive enough to assess temporal models and include in our analysis the cross-correlation between the ground truth and the imputed signal. We show that Shi-VAE achieves the best performance in terms of using both metrics, with lower computational complexity than the GP-VAE model, which is the state-of-the-art method for medical records., Comment: Accepted for publication in IEEE Journal of Biomedical and Health Informatics (JBHI)

Language models (LM) have grown with non-stop in the last decade, from sequence-to-sequence architectures to the state-of-the-art and utter attention-based Transformers. In this work, we demonstrate how the inclusion of deep generative models within BERT can bring more versatile models, able to impute missing/noisy words with richer text or even improve BLEU score. More precisely, we use a Gaussian Mixture Variational Autoencoder (GMVAE) as a regularizer layer and prove its effectiveness not only in Transformers but also in the most relevant encoder-decoder based LM, seq2seq with and without attention.

Article number 9438636 In this paper we address the problem of offline handwritten text recognition (HTR) in historical documents when few labeled samples are available and some of them contain errors in the train set. Our three main contributions are: first, we analyze how to perform transfer learning (TL) from a massive database to a smaller historical database, analyzing which layers of the model need fine-tuning. Second, we analyze methods to efficiently combine TL and data augmentation (DA). Finally, we propose an algorithm to mitigate the effects of incorrect labeling in the training set. The methods are analyzed over the ICFHR 2018 competition database, Washington and Parzival. Combining all these techniques, we demonstrate a remarkable reduction of CER (up to 6 percentage points in some cases) in the test set with little complexity overhead.

Generalized low-density parity-check (GLDPC) codes, where the single parity-check (SPC) nodes are replaced by generalized constraint (GC) nodes, are known to offer a reduced gap to capacity when compared with conventional LDPC codes, while also maintaining linear growth of minimum distance. However, for certain classes of practical GLDPC codes, there remains a gap to capacity even when utilizing blockwise decoding algorithm at GC nodes. In this work, we propose to optimize the design of GLDPC codes where the GC nodes are decoded with a trellis-based bit-wise Bahl-Cocke-Jelinek- Raviv (BCJR) component decoding algorithm. We analyze the asymptotic threshold behavior of GLDPC codes and determine the optimal proportion of the GC nodes in the GLDPC Tanner graph.We show significant performance improvements compared to existing designs with the same order of decoding complexity.