Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318037
H. Saruwatari, H. Yamajo, T. Takatani, T. Nishikawa, K. Shikano
We propose a new two-stage blind separation and deconvolution algorithm for multiple-input multiple-output (MIMO)- FIR system driven by colored sound sources, in which a new single-input multiple-output (SIMO)-model-based ICA (SIMO-ICA) and blind multichannel inverse filtering are combined. SIMO-ICA can separate the mixed signals, not into monaural source signals but into SIMO-model-based signals from independent sources. After SIMO-ICA, a simple blind deconvolution technique for the SIMO model can be applied even when each source signal is temporally correlated. The simulation results reveal that the proposed algorithm can successfully achieve the separation and deconvolution for a convolutive mixture of speech.
{"title":"Blind separation and deconvolution of MIMO system driven by colored inputs using SIMO-model-based ICA with information-geometric learning","authors":"H. Saruwatari, H. Yamajo, T. Takatani, T. Nishikawa, K. Shikano","doi":"10.1109/NNSP.2003.1318037","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318037","url":null,"abstract":"We propose a new two-stage blind separation and deconvolution algorithm for multiple-input multiple-output (MIMO)- FIR system driven by colored sound sources, in which a new single-input multiple-output (SIMO)-model-based ICA (SIMO-ICA) and blind multichannel inverse filtering are combined. SIMO-ICA can separate the mixed signals, not into monaural source signals but into SIMO-model-based signals from independent sources. After SIMO-ICA, a simple blind deconvolution technique for the SIMO model can be applied even when each source signal is temporally correlated. The simulation results reveal that the proposed algorithm can successfully achieve the separation and deconvolution for a convolutive mixture of speech.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121139304","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318028
Samuel Kaski, J. Sinkkonen, Arto Klami
A generative distributional clustering model for continuous data is reviewed and methods for optimizing and regularizing it are introduced and compared. Based on pairs of auxiliary and primary data, the primary data space is partitioned into Voronoi regions that are maximally homogeneous in terms of auxiliary data. Then only variation in the primary data associated with variation in the auxiliary data influences the clusters. Because the whole primary space is partitioned, new samples can be easily clustered in terms of primary data alone. In experiments, the approach is shown to produce more homogeneous clusters than alternative methods. Two regularization methods are demonstrated to further improve the results: an entropy-type penalty for unequal cluster sizes, and the inclusion of a K-means component to the model. The latter can alternatively be interpreted as special kind of joint distribution modeling where the emphasis between discrimination and unsupervised modeling of primary data can be tuned.
{"title":"Regularized discriminative clustering","authors":"Samuel Kaski, J. Sinkkonen, Arto Klami","doi":"10.1109/NNSP.2003.1318028","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318028","url":null,"abstract":"A generative distributional clustering model for continuous data is reviewed and methods for optimizing and regularizing it are introduced and compared. Based on pairs of auxiliary and primary data, the primary data space is partitioned into Voronoi regions that are maximally homogeneous in terms of auxiliary data. Then only variation in the primary data associated with variation in the auxiliary data influences the clusters. Because the whole primary space is partitioned, new samples can be easily clustered in terms of primary data alone. In experiments, the approach is shown to produce more homogeneous clusters than alternative methods. Two regularization methods are demonstrated to further improve the results: an entropy-type penalty for unequal cluster sizes, and the inclusion of a K-means component to the model. The latter can alternatively be interpreted as special kind of joint distribution modeling where the emphasis between discrimination and unsupervised modeling of primary data can be tuned.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"17 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128079079","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318017
Keisuke Yamazaki, Sumio Watanabe
Hidden Markov models are now used in many fields, for example, speech recognition, natural language processing etc. However, the mathematical foundation of analysis for the models has not yet been constructed, since the HMMs are non-identifiable. In recent years, we have developed the algebraic geometrical method that allows us to analyze the non-regular and non-identifiable models. In this paper, we apply this method to the HMM and reveal the asymptotic order of its stochastic complexity in the mathematically rigorous way.
{"title":"Stochastic complexities of hidden Markov models","authors":"Keisuke Yamazaki, Sumio Watanabe","doi":"10.1109/NNSP.2003.1318017","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318017","url":null,"abstract":"Hidden Markov models are now used in many fields, for example, speech recognition, natural language processing etc. However, the mathematical foundation of analysis for the models has not yet been constructed, since the HMMs are non-identifiable. In recent years, we have developed the algebraic geometrical method that allows us to analyze the non-regular and non-identifiable models. In this paper, we apply this method to the HMM and reveal the asymptotic order of its stochastic complexity in the mathematically rigorous way.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"398 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131848121","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318046
M. Barnard, J. Odobez, Samy Bengio
The recognition of events within multi-modal data is a challenging problem. In this paper we focus on the recognition of events by using both audio and video data. We investigate the use of data fusion techniques in order to recognise these sequences within the framework of hidden Markov models (HMM) used to model audio and video data sequences. Specifically we look at the recognition of play and break sequences in football and the segmentation of football games based on these two events. Recognising relatively simple semantic events such as this is an important step towards full automatic indexing of such video material. These experiments were done using approximately 3 hours of data from two games of the Euro96 competition. We propose that modelling the audio and video streams separately for each sequence and fusing the decisions from each stream should yield an accurate and robust method of segmenting multi-modal data.
{"title":"Multi-modal audio-visual event recognition for football analysis","authors":"M. Barnard, J. Odobez, Samy Bengio","doi":"10.1109/NNSP.2003.1318046","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318046","url":null,"abstract":"The recognition of events within multi-modal data is a challenging problem. In this paper we focus on the recognition of events by using both audio and video data. We investigate the use of data fusion techniques in order to recognise these sequences within the framework of hidden Markov models (HMM) used to model audio and video data sequences. Specifically we look at the recognition of play and break sequences in football and the segmentation of football games based on these two events. Recognising relatively simple semantic events such as this is an important step towards full automatic indexing of such video material. These experiments were done using approximately 3 hours of data from two games of the Euro96 competition. We propose that modelling the audio and video streams separately for each sequence and fusing the decisions from each stream should yield an accurate and robust method of segmenting multi-modal data.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"146 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133104277","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318047
A. Heittmann, U. Ramacher
The feature extraction and detection in visual scenes set up the basis for robust image processing and scene analysis. While the receptive fields of simple cells in the visual cortex are modeled by Gabor functions, simple cells are commonly treated as linear filters. In this paper, we demonstrate how the non-linear operations on pulses like correlation, synchronization and detection of decorrelation can be used for implementation of feature detectors. Using essentially two data-driven adaption rules dependent on dendritic currents and to membrane potentials, linear detection of intensity gradients can be realized. As a technical application, a feature detector sensitive to orientation is presented.
{"title":"Correlation-based feature detection using pulsed neural networks","authors":"A. Heittmann, U. Ramacher","doi":"10.1109/NNSP.2003.1318047","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318047","url":null,"abstract":"The feature extraction and detection in visual scenes set up the basis for robust image processing and scene analysis. While the receptive fields of simple cells in the visual cortex are modeled by Gabor functions, simple cells are commonly treated as linear filters. In this paper, we demonstrate how the non-linear operations on pulses like correlation, synchronization and detection of decorrelation can be used for implementation of feature detectors. Using essentially two data-driven adaption rules dependent on dendritic currents and to membrane potentials, linear detection of intensity gradients can be realized. As a technical application, a feature detector sensitive to orientation is presented.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114622496","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318007
L. Andrade, E. Manolakos
A physical understanding of mobility pattern differences among DNA fragments of different length drives the formulation of a complete statistical model for mobility shifts correction (MSC). Algorithms are then developed that compute mobility shifts and correct the raw trace, leaving significant peaks aligned i.e. in correct order and centered with respect to their neighbors. The fully automated MSC method we develop is shown to improve substantially base-calling performance in the first 100-200 bp of DNA sequencing reads while it does not require any dye chemistry specific calibration procedure. This is very encouraging since it is known that lack of adequate MSC is the main source of base-calling errors in the early part of a DNA sequencing read.
{"title":"Automatic estimation of mobility shift coefficients in DNA chromatograms","authors":"L. Andrade, E. Manolakos","doi":"10.1109/NNSP.2003.1318007","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318007","url":null,"abstract":"A physical understanding of mobility pattern differences among DNA fragments of different length drives the formulation of a complete statistical model for mobility shifts correction (MSC). Algorithms are then developed that compute mobility shifts and correct the raw trace, leaving significant peaks aligned i.e. in correct order and centered with respect to their neighbors. The fully automated MSC method we develop is shown to improve substantially base-calling performance in the first 100-200 bp of DNA sequencing reads while it does not require any dye chemistry specific calibration procedure. This is very encouraging since it is known that lack of adequate MSC is the main source of base-calling errors in the early part of a DNA sequencing read.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"92 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116338107","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318010
L. Gross-Colzy, R. Frouin
Satellite ocean-color algorithms generally use aerosol-mixture models to estimate and remove the atmospheric contribution to the measured signal. These models, based on aerosol samples, may or may not be realistic. To investigate the adequacy of the models and ultimately to improve atmospheric correction, we analyze atmospheric optics data collected by the aerosol robotic network project under a wide range of aerosol conditions at coastal and island sites. Using non-supervised classification techniques (probabilistic self-organized mapping), we determine the distribution of retrieved aerosol properties of the total atmospheric column, i.e., the volume size distribution function and the refractive index. The centers of the PRSOM neurons may be used as new aerosols models in radiative transfer algorithms.
{"title":"Self-organized mapping of aerosol mixtures at aeronet coastal and island sites","authors":"L. Gross-Colzy, R. Frouin","doi":"10.1109/NNSP.2003.1318010","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318010","url":null,"abstract":"Satellite ocean-color algorithms generally use aerosol-mixture models to estimate and remove the atmospheric contribution to the measured signal. These models, based on aerosol samples, may or may not be realistic. To investigate the adequacy of the models and ultimately to improve atmospheric correction, we analyze atmospheric optics data collected by the aerosol robotic network project under a wide range of aerosol conditions at coastal and island sites. Using non-supervised classification techniques (probabilistic self-organized mapping), we determine the distribution of retrieved aerosol properties of the total atmospheric column, i.e., the volume size distribution function and the refractive index. The centers of the PRSOM neurons may be used as new aerosols models in radiative transfer algorithms.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"33 1-2 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116717139","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318051
Chun-fu Lin, Sheng-de Wang
Fuzzy support vector machines (FSVMs) provide a method to classify data with noises or outliers. Each data point is associated with a fuzzy membership that can reflect their relative degrees as meaningful data. In this paper, we investigate and compare two strategies of automatically setting the fuzzy memberships of data points. It makes the usage of FSVMs easier in the application of reducing the effects of noises or outliers. The experiments show that the generalization error of FSVMs is comparable to other methods on benchmark datasets.
{"title":"Training algorithms for fuzzy support vector machines with noisy data","authors":"Chun-fu Lin, Sheng-de Wang","doi":"10.1109/NNSP.2003.1318051","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318051","url":null,"abstract":"Fuzzy support vector machines (FSVMs) provide a method to classify data with noises or outliers. Each data point is associated with a fuzzy membership that can reflect their relative degrees as meaningful data. In this paper, we investigate and compare two strategies of automatically setting the fuzzy memberships of data points. It makes the usage of FSVMs easier in the application of reducing the effects of noises or outliers. The experiments show that the generalization error of FSVMs is comparable to other methods on benchmark datasets.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"24 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125437847","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318015
Bruno Pelletier
In this paper, a neural network field over a subset /spl Xi/ of a metric space and a corresponding stochastic learning algorithm are introduced. A neural network field is a neural network, the parameters of which are functions of other variables, being valued in /spl Xi/. Neural network fields are mostly dedicated to the problem of approximating a parametrized function or, more generally, to the problem of approximating a function field. Typical examples of this kind of problem may be found in the context of geophysical sciences, where the observed data depends on two or three angular variables characterizing the data acquisition process. Neural network fields also offers interesting perspectives within the field of parametric nonlinear modeling techniques.
{"title":"Neural network fields","authors":"Bruno Pelletier","doi":"10.1109/NNSP.2003.1318015","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318015","url":null,"abstract":"In this paper, a neural network field over a subset /spl Xi/ of a metric space and a corresponding stochastic learning algorithm are introduced. A neural network field is a neural network, the parameters of which are functions of other variables, being valued in /spl Xi/. Neural network fields are mostly dedicated to the problem of approximating a parametrized function or, more generally, to the problem of approximating a function field. Typical examples of this kind of problem may be found in the context of geophysical sciences, where the observed data depends on two or three angular variables characterizing the data acquisition process. Neural network fields also offers interesting perspectives within the field of parametric nonlinear modeling techniques.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125834943","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2003-09-17DOI: 10.1109/NNSP.2003.1318004
C. Benros, A. G. Brevern, S. Hazout
Predicting protein structure from amino acid sequence is one of the main challenges of genomics. Various computational methods have been developed during the last decade to reach this goal. However, the problem of structure prediction remains difficult. Before facing this complex problem, our goal is to focus on the accurate analysis of protein structures at a local level. In our study, we present an approach called "hybrid protein model" (HPM) which uses a training procedure similar to the one of the self-organizing maps. It allows the compression of a non-redundant protein structure databank into a library of overlapping 3D structural fragments. The "hybrid protein model" carries out a multiple alignment of structural fragments. We present in this study an improvement of this strategy by introducing gaps in the local structures, and a sensitivity study of the training according to the control parameters. The library obtained is composed of a finite number of structural classes, each class including fragments sharing similar local structures. These classes are representative of the structural motifs found in the protein structures from the databank. Thus, this library constitutes an efficient tool for determining structural similarities between proteins and especially for predicting the local protein structure from the amino acid sequence.
{"title":"Hybrid protein model (HPM): a method for building a library of overlapping local structural prototypes. Sensitivity study and improvements of the training","authors":"C. Benros, A. G. Brevern, S. Hazout","doi":"10.1109/NNSP.2003.1318004","DOIUrl":"https://doi.org/10.1109/NNSP.2003.1318004","url":null,"abstract":"Predicting protein structure from amino acid sequence is one of the main challenges of genomics. Various computational methods have been developed during the last decade to reach this goal. However, the problem of structure prediction remains difficult. Before facing this complex problem, our goal is to focus on the accurate analysis of protein structures at a local level. In our study, we present an approach called \"hybrid protein model\" (HPM) which uses a training procedure similar to the one of the self-organizing maps. It allows the compression of a non-redundant protein structure databank into a library of overlapping 3D structural fragments. The \"hybrid protein model\" carries out a multiple alignment of structural fragments. We present in this study an improvement of this strategy by introducing gaps in the local structures, and a sensitivity study of the training according to the control parameters. The library obtained is composed of a finite number of structural classes, each class including fragments sharing similar local structures. These classes are representative of the structural motifs found in the protein structures from the databank. Thus, this library constitutes an efficient tool for determining structural similarities between proteins and especially for predicting the local protein structure from the amino acid sequence.","PeriodicalId":315958,"journal":{"name":"2003 IEEE XIII Workshop on Neural Networks for Signal Processing (IEEE Cat. No.03TH8718)","volume":"142 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2003-09-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121556904","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：