Human beings are a source of many signals that encode information related to the health status of the person, its intentions, the status of diverse biological process in his body, the information he is trying to communicate, etc. In general these signals are of nonlinear and nonstationary nature, and the information is hidden and coded in a complex way which is not clearly exposed in the raw signal. In many cases, the information conveyed in those signals can be interpreted rather easily by other human beings. For example, a listener can understand the meaning of a speech sentence and the emotional state of the speaker. As another example, an expert can determine the presence or absence of apnea events by watching a polysomnogram of a sleeping patient. Nevertheless, when one tries to replicate this characteristic in a computer, the nonlinear and nonstationary characteristics pose a serious problem for an automatic algorithm. Moreover, there are signals that carry interesting information that at the present day cannot be interpreted by other humans. For example, the electroencephalographic record obtaining at the same time a person is imagining to pronounce a word, carry information regarding the word that was imagined, but it is not possible for a human observer to extract this information.In real cases, most of the signals will be affected by different kinds of noise, artifacts, mixing with other non-interesting signals, etc. This aspect causes a strong problem as most systems are built under clean, ideal conditions. In this way we need advanced machine learning and signal processing methods to produce estimation of the clean signals. In the last years several advanced signal processing methods have been developed which try to improve the capabilities of information extraction. Some of these techniques are sparse representation, nonnegative matrix factorization, blind source separation, independent component analysis, regularization and inverse problems.The objectives of this research line are: -To develop advances algorithms that can enhance the quality of the required signals (this includes blind signal separation, denoising, enhancing, dereverberation). -To produce advanced methods to extract useful information from the signals under analysis (this includes sparse representations, nonnegative matrix factorization, discriminant basis construction, among others). -To develop advanced machine learning methods which can be used to detect the presence or absence of some characteristic wave, pattern or pathology (this includes the use of deep learning, recurrent neural networks, deep belief networks, spiked neural networks, among others) . We apply these methods to a variety of signals, including speech, electrocardiogram signals, electroencephalogram signals, pulse oximetry signals, arterial pressure signals, among others, and develop a variety of applications like emotion recognition, speech recognition, speech denoising and dereverberation, detection of apneas and hypopneas, prediction of the presence of the sleep apnea-hypopneas syndrome and other related applications
Leonardo Rufiner, Leandro Di Persia, Roman Rolon, Ivan Gareis, Jose Chelotti, Sebastian Vanrell
Team members
Leonardo Rufiner
Leandro Di Persia
Roman Rolon
Ivan Gareis
Jose Chelotti
Sebastian Vanrell
Web demos
mirDNN: A convolutional deep residual neural network for prediction of pre-miRNAs in genome-wide data.
DL4papers: A deep learning approach for the automatic interpretation of scientific articles.
PhDSeeker: PhDSeeker is a bio-inspired algorithm for synthesizing linear and branched metabolic pathways.
Clustermatch: Clustermatch is an efficient clustering method able to process highly diverse datasest.
ELM-GRNNminer: Extreme learning machines for reverse engineering of gene regulatory networks from expression time series.
miRNAss: Semi-supervised method for microRNA prediction with few labelled samples.
BioDataFusion: A novel approach for highly-diverse multi-omics data fusion.
deepSOM: A novel and effective way of approaching high class-imbalance in pre-miRNA prediction.
miRNA: supervised vs unsupervised Compares a supervised vs an unsupervised approach for pre-miRNA prediction in model genomes.
gamma-AM: A novel method for inferring biological function for a set of genes with previously unknown function.
divControl: A novel method to smoothly control the diversity of a cluster ensemble.
miRNA-SOM: miRNA-SOM is a tool for the discovery of pre-miRNA in the E. multilocularis genome.
EvoMS: An evolutionary tool for finding novel metabolic pathways linking compounds through feasible reactions.
miRNAfe: A comprehensive tool providing almost all state-of-the-art RNA feature extraction methods used today.
miRNAfe: full miRNAfe full is an advanced tool to extract features from RNA sequences, providing almost all state-of-the-art feature extraction methods published today.
omeSOM full: *omeSOM is a tool designed to give support to the data mining task of metabolic and transcriptional datasets derived from different databases.
bSOM lite: Simple demo of bSOM, an algorithm for improve clustering with information from metabolic pathways.
Selected publications
Complexity measures of the mature miRNA for improving pre-miRNAs prediction <img class=”doc” src=”http://sinc.unl.edu.ar/sinc/sinc-publications/images/doi_small.png”











