Pub Date : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553390
Ayoub Hajlaoui, M. Chetouani, S. Essid
Electroencephalographic sensors have proven to be promising for emotion recognition. Our study focuses on the recognition of valence and arousal levels using such sensors. Usually, ad hoc features are extracted for such recognition tasks. In this paper, we rely on automatic feature learning techniques instead. Our main contribution is the use of Group Nonnegative Matrix Factorization in a multi-task fashion, where we exploit both valence and arousal labels to control valence-related and arousal-related feature learning. Applying this method on HCI MAHNOB and EMOEEG, two databases where emotions are elicited by means of audiovisual stimuli and performing binary inter-session classification of valence labels, we obtain significant improvement of valence classification Fl scores in comparison to baseline frequency-band power features computed on predefined frequency bands. The valence classification F1 score is improved from 0.56 to 0.69 in the case of HCI MAHNOB, and from 0.56 to 0.59 in the case of EMOEEG.
{"title":"Multi-task Feature Learning for EEG-based Emotion Recognition Using Group Nonnegative Matrix Factorization","authors":"Ayoub Hajlaoui, M. Chetouani, S. Essid","doi":"10.23919/EUSIPCO.2018.8553390","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553390","url":null,"abstract":"Electroencephalographic sensors have proven to be promising for emotion recognition. Our study focuses on the recognition of valence and arousal levels using such sensors. Usually, ad hoc features are extracted for such recognition tasks. In this paper, we rely on automatic feature learning techniques instead. Our main contribution is the use of Group Nonnegative Matrix Factorization in a multi-task fashion, where we exploit both valence and arousal labels to control valence-related and arousal-related feature learning. Applying this method on HCI MAHNOB and EMOEEG, two databases where emotions are elicited by means of audiovisual stimuli and performing binary inter-session classification of valence labels, we obtain significant improvement of valence classification Fl scores in comparison to baseline frequency-band power features computed on predefined frequency bands. The valence classification F1 score is improved from 0.56 to 0.69 in the case of HCI MAHNOB, and from 0.56 to 0.59 in the case of EMOEEG.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"54 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"126652542","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553254
A. Bouchain, A. Vercoutter, J. Picheral, A. Talon
Blades vibrations must be measured in operations to validate blade design. Tip-timing is one of the classical measurement methods but its main drawback is the generation of sub-sampled and non-uniform sampled signals. This paper presents a new sparse method for tip-timing spectral analysis that makes use of engine rotation variations. Assuming that blade vibration signals yield to line spectra, a sparse signal model is introduced as a linear system. The solution to the problem is obtained by ADMM (Alternating Direction Method of Multipliers) with a $p^{1}$ -regularization. Results for simulated and real signals are given to illustrate the efficiency of this method. The main advantages of the proposed method are to provide a fast solution and to take into account the variations of the rotation speed. Results show that this approach reduces frequency aliasings caused by the low sampling frequency of the measured signals.
{"title":"Sparse Method for Tip-Timing Signals Analysis with Non Stationary Engine Rotation Frequency","authors":"A. Bouchain, A. Vercoutter, J. Picheral, A. Talon","doi":"10.23919/EUSIPCO.2018.8553254","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553254","url":null,"abstract":"Blades vibrations must be measured in operations to validate blade design. Tip-timing is one of the classical measurement methods but its main drawback is the generation of sub-sampled and non-uniform sampled signals. This paper presents a new sparse method for tip-timing spectral analysis that makes use of engine rotation variations. Assuming that blade vibration signals yield to line spectra, a sparse signal model is introduced as a linear system. The solution to the problem is obtained by ADMM (Alternating Direction Method of Multipliers) with a $p^{1}$ -regularization. Results for simulated and real signals are given to illustrate the efficiency of this method. The main advantages of the proposed method are to provide a fast solution and to take into account the variations of the rotation speed. Results show that this approach reduces frequency aliasings caused by the low sampling frequency of the measured signals.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"490 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115535039","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553100
K. Yuan, Bicheng Ying, A. H. Sayed
This work develops a fully decentralized variance-reduced learning algorithm for multi-agent networks where nodes store and process the data locally and are only allowed to communicate with their immediate neighbors. In the proposed algorithm, there is no need for a central or master unit while the objective is to enable the dispersed nodes to learn the exact global model despite their limited localized interactions. The resulting algorithm is shown to have low memory requirement, guaranteed linear convergence, robustness to failure of links or nodes and scalability to the network size. Moreover, the decentralized nature of the solution makes large-scale machine learning problems more tractable and also scalable since data is stored and processed locally at the nodes.
{"title":"Efficient Variance-Reduced Learning Over Multi-Agent Networks","authors":"K. Yuan, Bicheng Ying, A. H. Sayed","doi":"10.23919/EUSIPCO.2018.8553100","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553100","url":null,"abstract":"This work develops a fully decentralized variance-reduced learning algorithm for multi-agent networks where nodes store and process the data locally and are only allowed to communicate with their immediate neighbors. In the proposed algorithm, there is no need for a central or master unit while the objective is to enable the dispersed nodes to learn the exact global model despite their limited localized interactions. The resulting algorithm is shown to have low memory requirement, guaranteed linear convergence, robustness to failure of links or nodes and scalability to the network size. Moreover, the decentralized nature of the solution makes large-scale machine learning problems more tractable and also scalable since data is stored and processed locally at the nodes.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"112 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115985980","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553354
Jessica Sena, Jesimon Barreto Santos, W. R. Schwartz
Sensor-based Human Activity Recognition (HAR) provides valuable knowledge to many areas. Recently, wearable devices have gained space as a relevant source of data. However, there are two issues: large number of heterogeneous sensors available and the temporal nature of the sensor data. To handle those issues, we propose a multimodal approach that processes each sensor separately and, through an ensemble of Deep Convolution Neural Networks (DCNN), extracts information from multiple temporal scales of the sensor data. In this ensemble, we use a convolutional kernel with a different height for each DCNN. Considering that the number of rows in the sensor data reflects the data captured over time, each kernel height reflects a temporal scale from which we can extract patterns. Consequently, our approach is able to extract from simple movement patterns such as a wrist twist when picking up a spoon to complex movements such as the human gait. This multimodal and multitemporal approach outperforms previous state-of-the-art works in seven important datasets using two different protocols. In addition, we demonstrate that the use of our proposed set of kernels improves sensor-based HAR in another multi-kernel approach, the widely employed inception network.
{"title":"Multiscale DCNN Ensemble Applied to Human Activity Recognition Based on Wearable Sensors","authors":"Jessica Sena, Jesimon Barreto Santos, W. R. Schwartz","doi":"10.23919/EUSIPCO.2018.8553354","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553354","url":null,"abstract":"Sensor-based Human Activity Recognition (HAR) provides valuable knowledge to many areas. Recently, wearable devices have gained space as a relevant source of data. However, there are two issues: large number of heterogeneous sensors available and the temporal nature of the sensor data. To handle those issues, we propose a multimodal approach that processes each sensor separately and, through an ensemble of Deep Convolution Neural Networks (DCNN), extracts information from multiple temporal scales of the sensor data. In this ensemble, we use a convolutional kernel with a different height for each DCNN. Considering that the number of rows in the sensor data reflects the data captured over time, each kernel height reflects a temporal scale from which we can extract patterns. Consequently, our approach is able to extract from simple movement patterns such as a wrist twist when picking up a spoon to complex movements such as the human gait. This multimodal and multitemporal approach outperforms previous state-of-the-art works in seven important datasets using two different protocols. In addition, we demonstrate that the use of our proposed set of kernels improves sensor-based HAR in another multi-kernel approach, the widely employed inception network.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"47 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116068641","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553295
M. Kharouf, Tabea Rebafka, Nataliya Sokolovska
This paper addresses the problem of dimension reduction of noisy data, more precisely the challenge to determine the dimension of the subspace where the observed signal lives in. Based on results from random matrix theory, two novel estimators of the signal dimension are proposed in this paper. Consistency of the estimators is proved in the modern asymptotic regime, where the number of parameters grows proportionally with the sample size. Experimental results show that the novel estimators are robust to noise and, moreover, they give highly accurate results in settings where standard methods fail. We apply the novel dimension estimators to several life sciences benchmarks in the context of classification, and illustrate the improvements achieved by the new methods compared to the state-of-the-art approaches.
{"title":"Consistent Spectral Methods for Dimensionality Reduction","authors":"M. Kharouf, Tabea Rebafka, Nataliya Sokolovska","doi":"10.23919/EUSIPCO.2018.8553295","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553295","url":null,"abstract":"This paper addresses the problem of dimension reduction of noisy data, more precisely the challenge to determine the dimension of the subspace where the observed signal lives in. Based on results from random matrix theory, two novel estimators of the signal dimension are proposed in this paper. Consistency of the estimators is proved in the modern asymptotic regime, where the number of parameters grows proportionally with the sample size. Experimental results show that the novel estimators are robust to noise and, moreover, they give highly accurate results in settings where standard methods fail. We apply the novel dimension estimators to several life sciences benchmarks in the context of classification, and illustrate the improvements achieved by the new methods compared to the state-of-the-art approaches.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"51 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122380613","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553502
Sean M. Kennedy, J. Roth, James W. Scrofani
In this paper, we propose a novel method for embedding one-dimensional, periodic time-series data into higher-dimensional topological spaces to support robust recovery of signal features via topological data analysis under noisy sampling conditions. Our method can be considered an extension of the popular time delay embedding method to a larger class of linear operators. To provide evidence for the viability of this method, we analyze the simple case of sinusoidal data in three steps. First, we discuss some of the drawbacks of the time delay embedding framework in the context of periodic, sinusoidal data. Next, we show analytically that using the Hilbert transform as an alternative embedding function for sinusoidal data overcomes these drawbacks. Finally, we provide empirical evidence of the viability of the Hilbert transform as an embedding function when the parameters of the sinusoidal data vary over time.
{"title":"A Novel Method for Topological Embedding of Time-Series Data","authors":"Sean M. Kennedy, J. Roth, James W. Scrofani","doi":"10.23919/EUSIPCO.2018.8553502","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553502","url":null,"abstract":"In this paper, we propose a novel method for embedding one-dimensional, periodic time-series data into higher-dimensional topological spaces to support robust recovery of signal features via topological data analysis under noisy sampling conditions. Our method can be considered an extension of the popular time delay embedding method to a larger class of linear operators. To provide evidence for the viability of this method, we analyze the simple case of sinusoidal data in three steps. First, we discuss some of the drawbacks of the time delay embedding framework in the context of periodic, sinusoidal data. Next, we show analytically that using the Hilbert transform as an alternative embedding function for sinusoidal data overcomes these drawbacks. Finally, we provide empirical evidence of the viability of the Hilbert transform as an embedding function when the parameters of the sinusoidal data vary over time.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"61 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"122999181","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553466
Yassine Zniyed, R. Boyer, A. Almeida, G. Favier
The canonical polyadic decomposition (CPD) is one of the most popular tensor-based analysis tools due to its usefulness in numerous fields of application. The Q-order CPD is parametrized by $Q$ matrices also called factors which have to be recovered. The factors estimation is usually carried out by means of the alternating least squares (ALS) algorithm. In the context of multi-modal big data analysis, i.e., large order $(Q)$ and dimensions, the ALS algorithm has two main drawbacks. Firstly, its convergence is generally slow and may fail, in particular for large values of $Q$, and secondly it is highly time consuming. In this paper, it is proved that a Q-order CPD of rank-R is equivalent to a train of $Q$ 3-order CPD(s) of rank-R. In other words, each tensor train (TT)-core admits a 3-order CPD of rank-R. Based on the structure of the TT-cores, a new dimensionality reduction and factor retrieval scheme is derived. The proposed method has a better robustness to noise with a smaller computational cost than the ALS algorithm.
{"title":"High-Order CPD Estimation with Dimensionality Reduction Using a Tensor Train Model","authors":"Yassine Zniyed, R. Boyer, A. Almeida, G. Favier","doi":"10.23919/EUSIPCO.2018.8553466","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553466","url":null,"abstract":"The canonical polyadic decomposition (CPD) is one of the most popular tensor-based analysis tools due to its usefulness in numerous fields of application. The Q-order CPD is parametrized by $Q$ matrices also called factors which have to be recovered. The factors estimation is usually carried out by means of the alternating least squares (ALS) algorithm. In the context of multi-modal big data analysis, i.e., large order $(Q)$ and dimensions, the ALS algorithm has two main drawbacks. Firstly, its convergence is generally slow and may fail, in particular for large values of $Q$, and secondly it is highly time consuming. In this paper, it is proved that a Q-order CPD of rank-R is equivalent to a train of $Q$ 3-order CPD(s) of rank-R. In other words, each tensor train (TT)-core admits a 3-order CPD of rank-R. Based on the structure of the TT-cores, a new dimensionality reduction and factor retrieval scheme is derived. The proposed method has a better robustness to noise with a smaller computational cost than the ALS algorithm.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"35 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123020158","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553039
Mathew Shaji Kavalekalam, J. Nielsen, Liming Shi, M. G. Christensen, J. Boldt
In this paper, we propose a speech enhancement method based on non-negative matrix factorization (NMF) techniques. NMF techniques allow us to approximate the power spectral density (PSD) of the noisy signal as a weighted linear combination of trained speech and noise basis vectors arranged as the columns of a matrix. In this work, we propose to use basis vectors that are parameterised by autoregressive (AR) coefficients. Parametric representation of the spectral basis is beneficial as it can encompass the signal characteristics like, e.g. the speech production model. It is observed that the parametric representation of basis vectors is beneficial while performing online speech enhancement in low delay scenarios.
{"title":"Online Parametric NMF for Speech Enhancement","authors":"Mathew Shaji Kavalekalam, J. Nielsen, Liming Shi, M. G. Christensen, J. Boldt","doi":"10.23919/EUSIPCO.2018.8553039","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553039","url":null,"abstract":"In this paper, we propose a speech enhancement method based on non-negative matrix factorization (NMF) techniques. NMF techniques allow us to approximate the power spectral density (PSD) of the noisy signal as a weighted linear combination of trained speech and noise basis vectors arranged as the columns of a matrix. In this work, we propose to use basis vectors that are parameterised by autoregressive (AR) coefficients. Parametric representation of the spectral basis is beneficial as it can encompass the signal characteristics like, e.g. the speech production model. It is observed that the parametric representation of basis vectors is beneficial while performing online speech enhancement in low delay scenarios.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"12 7","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114128467","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553565
Sanna Wager, Minje Kim
We propose a regularized nonnegative tensor factorization (NTF) model for multi-channel speech derestriction that incorporates prior knowledge about clean speech. The approach models the problem as recovering a signal convolved with different room impulse responses, allowing the dereverberation problem to benefit from microphone arrays. The factorization learns both individual reverberation filters and channel-specific delays, which makes it possible to employ an ad-hoc microphone array with heterogeneous sensors (such as multi-channel recordings by a crowd) even if they are not synchronized. We integrate two prior-knowledge regularization schemes to increase the stability of dereverberation performance. First, a Nonnegative Matrix Factorization (NMF) inner routine is introduced to inform the original NTF problem of the pre-trained clean speech basis vectors, so that the optimization process can focus on estimating their activations rather than the whole clean speech spectra. Second, the NMF activation matrix is further regularized to take on characteristics of dry signals using sparsity and smoothness constraints. Empirical dereverberation results on different simulated reverberation setups show that the prior-knowledge regularization schemes improve both recovered sound quality and speech intelligibility compared to a baseline NTF approach.
{"title":"Collaborative Speech Dereverberation: Regularized Tensor Factorization for Crowdsourced Multi-Channel Recordings","authors":"Sanna Wager, Minje Kim","doi":"10.23919/EUSIPCO.2018.8553565","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553565","url":null,"abstract":"We propose a regularized nonnegative tensor factorization (NTF) model for multi-channel speech derestriction that incorporates prior knowledge about clean speech. The approach models the problem as recovering a signal convolved with different room impulse responses, allowing the dereverberation problem to benefit from microphone arrays. The factorization learns both individual reverberation filters and channel-specific delays, which makes it possible to employ an ad-hoc microphone array with heterogeneous sensors (such as multi-channel recordings by a crowd) even if they are not synchronized. We integrate two prior-knowledge regularization schemes to increase the stability of dereverberation performance. First, a Nonnegative Matrix Factorization (NMF) inner routine is introduced to inform the original NTF problem of the pre-trained clean speech basis vectors, so that the optimization process can focus on estimating their activations rather than the whole clean speech spectra. Second, the NMF activation matrix is further regularized to take on characteristics of dry signals using sparsity and smoothness constraints. Empirical dereverberation results on different simulated reverberation setups show that the prior-knowledge regularization schemes improve both recovered sound quality and speech intelligibility compared to a baseline NTF approach.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"14 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114132204","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 : 2018-09-01DOI: 10.23919/EUSIPCO.2018.8553034
Romain Couillet, Zhenyu Liao, Xiaoyi Mai
This article discusses the asymptotic performance of classical machine learning classification methods (from discriminant analysis to neural networks) for simultaneously large and numerous Gaussian mixture modelled data. We first provide theoretical bounds on the minimally discriminable class means and covariances under an oracle setting, which are then compared to recent theoretical findings on the performance of machine learning. Non-obvious phenomena are discussed, among which surprising phase transitions in the optimal performance rates for specific hyperparameter settings.
{"title":"Classification Asymptotics in the Random Matrix Regime","authors":"Romain Couillet, Zhenyu Liao, Xiaoyi Mai","doi":"10.23919/EUSIPCO.2018.8553034","DOIUrl":"https://doi.org/10.23919/EUSIPCO.2018.8553034","url":null,"abstract":"This article discusses the asymptotic performance of classical machine learning classification methods (from discriminant analysis to neural networks) for simultaneously large and numerous Gaussian mixture modelled data. We first provide theoretical bounds on the minimally discriminable class means and covariances under an oracle setting, which are then compared to recent theoretical findings on the performance of machine learning. Non-obvious phenomena are discussed, among which surprising phase transitions in the optimal performance rates for specific hyperparameter settings.","PeriodicalId":303069,"journal":{"name":"2018 26th European Signal Processing Conference (EUSIPCO)","volume":"93 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2018-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"114599429","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：