PyTorchKR Papers

최신 논문

총 11,809편
태그·기간 등 상세 필터

카테고리 탐색

로그인 후 이 태그 구독

arXiv:2006.07063

Privacy Against Adversarial Classification in Cyber-Physical Systems

For a class of Cyber-Physical Systems (CPSs), we address the problem of performing computations over the cloud without revealing private information about the structure and operation of the system. We model CPSs as a collection of input-output dynamical systems (the system operation modes). Depending on the mode the system is operating on, the output trajectory is generated by one of these systems in response to driving inputs. Output measurements and driving inputs are sent to the cloud for processing purposes. We capture this "processing" through some function (of the input-output trajectory) that we require the cloud to compute accurately - referred here as the trajectory utility. However, for privacy reasons, we would like to keep the mode private, i.e., we do not want the cloud to correctly identify what mode of the CPS produced a given trajectory. To this end, we distort trajectories before transmission and send the corrupted data to the cloud. We provide mathematical tools (based on output-regulation techniques) to properly design distorting mechanisms so that: 1) the original and distorted trajectories lead to the same utility; and the distorted data leads the cloud to misclassify the mode.

arXiv:2006.06682

Power System Disturbance Classification with Online Event-Driven Neuromorphic Computing

Accurate online classification of disturbance events in a transmission network is an important part of wide-area monitoring. Although many conventional machine learning techniques are very successful in classifying events, they rely on extracting information from PMU data at control centers and processing them through CPU/GPUs, which are highly inefficient in terms of energy consumption. To solve this challenge without compromising accuracy, this paper presents a novel methodology based on event-driven neuromorphic computing architecture for classification of power system disturbances. A Spiking Neural Network (SNN)-based computing framework is proposed, which exploits sparsity in disturbances and promotes local event driven operation for unsupervised learning and inference from incoming data. Spatio-temporal information of PMU signals is first extracted and encoded into spike trains and classification is achieved with SNN-based supervised and unsupervised learning framework. Moreover, a QR decomposition-based selection technique is proposed to identify signals participating in the low rank subspace of multiple disturbance events. Performance of the proposed method is validated on data collected from a 16-machine, 5-area New England-New York system.

arXiv:2005.14646

The INESC-ID Multi-Modal System for the ADReSS 2020 Challenge

This paper describes a multi-modal approach for the automatic detection of Alzheimer's disease proposed in the context of the INESC-ID Human Language Technology Laboratory participation in the ADReSS 2020 challenge. Our classification framework takes advantage of both acoustic and textual feature embeddings, which are extracted independently and later combined. Speech signals are encoded into acoustic features using DNN speaker embeddings extracted from pre-trained models. For textual input, contextual embedding vectors are first extracted using an English Bert model and then used either to directly compute sentence embeddings or to feed a bidirectional LSTM-RNNs with attention. Finally, an SVM classifier with linear kernel is used for the individual evaluation of the three systems. Our best system, based on the combination of linguistic and acoustic information, attained a classification accuracy of 81.25%. Results have shown the importance of linguistic features in the classification of Alzheimer's Disease, which outperforms the acoustic ones in terms of accuracy. Early stage features fusion did not provide additional improvements, confirming that the discriminant ability conveyed by speech in this case is smooth out by linguistic data.

arXiv:2005.11965

Automated MRI based pipeline for glioma segmentation and prediction of grade, IDH mutation and 1p19q co-deletion

In the WHO glioma classification guidelines grade, IDH mutation and 1p19q co-deletion play a central role as they are important markers for prognosis and optimal therapy planning. Therefore, we propose a fully automatic, MRI based, 3D pipeline for glioma segmentation and classification. The designed segmentation network was a 3D U-Net achieving an average whole tumor dice score of 90%. After segmentation, the 3D tumor ROI is extracted and fed into the multi-task classification network. The network was trained and evaluated on a large heterogeneous dataset of 628 patients, collected from The Cancer Imaging Archive and BraTS 2019 databases. Additionally, the network was validated on an independent dataset of 110 patients retrospectively acquired at the Ghent University Hospital (GUH). Classification AUC scores are 0.93, 0.94 and 0.82 on the TCIA test data and 0.94, 0.86 and 0.87 on the GUH data for grade, IDH and 1p19q status respectively.

arXiv:2004.11126

Leveraging Hardware-Impaired Out-of-Band Information Through Deep Neural Networks for Robust Wireless Device Classification

Wireless device classification techniques play a key role in promoting emerging wireless applications such as allowing spectrum regulatory agencies to enforce their access policies and enabling network administrators to control access and prevent impersonation attacks to their wireless networks. Leveraging spectrum distortions of transmitted RF signals, caused by transceiver hardware impairments created during manufacture and assembly stages, to provide device classification has been the focus of many recent works. These prior works essentially apply deep learning to extract features of the devices from their hardware-impaired signals and rely on feature variations across the devices to distinguish devices from one another. As technology advances, the manufacturing impairment variations across devices are becoming extremely insignificant, making these prior classification approaches inaccurate. This paper proposes a novel, deep learning-based technique that provides scalable and highly accurate classification of wireless devices, even when the devices exhibit insignificant variation across their hardware impairments and have the same hardware, protocol, and software configurations. The novelty of the proposed technique lies in leveraging both the {\em in-band} and {\em out-of-band} signal distortion information by oversampling the captured signals at the receiver and feeding IQ samples collected from the RF signals to a deep neural network for classification. Using a convolutional neural network (CNN) model, we show that our proposed technique, when applied to high-end, high-performance devices with minimally distorted hardware, doubles the device classification accuracy when compared to existing approaches.

arXiv:2004.08277

Learning Strategies for Radar Clutter Classification

In this paper, we address the problem of classifying clutter returns in order to partition them into statistically homogeneous subsets. The classification procedure relies on a model for the observables including latent variables that is solved by the expectation-maximization algorithm. The derivations are carried out by accounting for three different cases for the structure of the clutter covariance matrix. A preliminary performance analysis highlights that the proposed technique is a viable means to cluster clutter returns over the range.

arXiv:2003.13033

Mechanical classification of voice quality

While there is no a priori definition of good singing voices, we tend to make consistent evaluations of the quality of singing almost instantaneously. Such an instantaneous evaluation might be based on the sound spectrum that can be perceived in a short time. Here we devise a Bayesian algorithm that learns to evaluate the choral proficiency, musical scale, and gender of individual singers using the sound spectra of singing voices. In particular, the classification is performed on a set of sound spectral intensities, whose frequencies are selected by minimizing the Bayes risk. This optimization allows the algorithm to capture sound frequencies that are essential for each discrimination task, resulting in a good assessment performance. Experimental results revealed that a sound duration of about 0.1 sec is sufficient for determining the choral proficiency and gender of a singer. With a program constructed on this algorithm, everyone can evaluate choral voices of others and perform private vocal exercises.

arXiv:2003.12690

EEG Signal Classification using Variational Mode Decomposition

Epilepsy affects about 1% of the population every year, and is characterized by abnormal and sudden hyper-synchronous excitation of the neurons in the brain. The electroencephalogram(EEG) is the most widely used method to record brain signals and diagnose epilepsy and seizure cases. In this paper we use the method of Variational Mode Decomposition (VMD) in our analysis to classify seizure/seizure free signals. This technique uses variational non recursive mode decomposition, in comparison to other methods like Empirical Mode (EMD) and Hilbert-Huang transform which recursively decompose the signals, making them more susceptible to noise and sampling rate. VMD decomposes a signal into its components which are called principal modes. In our analysis, 4 features of the decomposed signals namely Renyi Entropy, second order difference plot (SODP), fourth order difference plot(FODP) and average amplitude are investigated, both individually and using a ranking methodology considering all 4 features at the same time. The SODP of decomposed signal modes is an elliptical structure. The 95% confidence ellipse area measured from the SODP of the decomposed signal modes has been used as a feature in order to discriminate seizure-free EEG signals from the epileptic seizure EEG signal. For the classification, a Multilayer Perceptron(MLP) with back propagation algorithm as the training method was used. A high percentage of accuracy was obtained when the features were used individually for classification and an even higher degree of accuracy was obtained when the ranking methodology was used.

arXiv:2003.12670

Objective Multi-variable Classification and Inference of Biological Neuronal Networks

Classification of biological neuron types and networks poses challenges to the full understanding of the brain's organisation and functioning. In this paper, we develop a novel objective classification model of biological neuronal types and networks based on the communication metrics of neurons. This presents advantages against the existing approaches since the mutual information or the delay between neurons obtained from spike trains are more abundant data compare to conventional morphological data. We firstly designed two open-access supporting computational platforms of various neuronal circuits from the Blue Brain Project realistic models, named Neurpy and Neurgen. Then we investigate how the concept of network tomography could be achieved with cortical neuronal circuits for morphological, topological and electrical classification of neurons. We extract the simulated data to many different classifiers (including SVM, Decision Trees, Random Forest, and Artificial Neuron Networks) classifying the specific cell type (and sub-group types) achieving accuracies of up to 70\%. Inference of biological network structures using network tomography reached up to 65\% of accuracy. We also analysed recall, precision and F1score of the classification of five layers, 25 cell m-types, and 14 cell e-types. Our research not only contributes to existing classification efforts but sets the road-map for future usage of cellular-scaled brain-machine interfaces for in-vivo objective classification of neurons as a sensing mechanism of the brain's structure.

arXiv:2003.10437

Rock Classification in Petrographic Thin Section Images Based on Concatenated Convolutional Neural Networks

Rock classification plays an important role in rock mechanics, petrology, mining engineering, magmatic processes, and numerous other fields pertaining to geosciences. This study proposes a concatenated convolutional neural network (Con-CNN) method for classifying the geologic rock type based on petrographic thin sections. Herein, plane polarized light (PPL) and crossed polarized light (XPL) were used to acquire thin section images as the fundamental data. After conducting the necessary pre-processing analyses, the PPL and XPL images as well as their comprehensive image (CI) were incorporated in three convolutional neural networks (CNNs) comprising the same structure for achieving a preliminary classification; these images were developed by employing the fused principal component analysis (PCA). Subsequently, the results of the CNNs were concatenated by using the maximum likelihood detection to obtain a comprehensive classification result. Finally, a statistical revision was applied to fix the misclassification due to the proportion difference of minerals that were similar in appearance. In this study, 13 types of 92 rock samples, 196 petrographic thin sections, 588 images, and 63504 image patches were fabricated for the training and validation of the Con-CNN. The five-folds cross validation shows that the method proposed provides an overall accuracy of 89.97%, which facilitates the automation of rock classification in petrographic thin sections.

arXiv:2003.09164

Acoustic Scene Classification using Audio Tagging

Acoustic scene classification systems using deep neural networks classify given recordings into pre-defined classes. In this study, we propose a novel scheme for acoustic scene classification which adopts an audio tagging system inspired by the human perception mechanism. When humans identify an acoustic scene, the existence of different sound events provides discriminative information which affects the judgement. The proposed framework mimics this mechanism using various approaches. Firstly, we employ three methods to concatenate tag vectors extracted using an audio tagging system with an intermediate hidden layer of an acoustic scene classification system. We also explore the multi-head attention on the feature map of an acoustic scene classification system using tag vectors. Experiments conducted on the detection and classification of acoustic scenes and events 2019 task 1-a dataset demonstrate the effectiveness of the proposed scheme. Concatenation and multi-head attention show a classification accuracy of 75.66 % and 75.58 %, respectively, compared to 73.63 % accuracy of the baseline. The system with the proposed two approaches combined demonstrates an accuracy of 76.75 %.

arXiv:2002.06391

Single-cell entropy to quantify the cellular transcription from single-cell RNA-seq data

We present the use of single-cell entropy (scEntropy) to measure the order of the cellular transcriptome profile from single-cell RNA-seq data, which leads to a method of unsupervised cell type classification through scEntropy followed by the Gaussian mixture model (scEGMM). scEntropy is straightforward in defining an intrinsic transcriptional state of a cell. scEGMM is a coherent method of cell type classification that includes no parameters and no clustering; however, it is comparable to existing machine learning-based methods in benchmarking studies and facilitates biological interpretation.

arXiv:2002.06279

A Comparison of Pooling Methods on LSTM Models for Rare Acoustic Event Classification

Acoustic event classification (AEC) and acoustic event detection (AED) refer to the task of detecting whether specific target events occur in audios. As long short-term memory (LSTM) leads to state-of-the-art results in various speech related tasks, it is employed as a popular solution for AEC as well. This paper focuses on investigating the dynamics of LSTM model on AEC tasks. It includes a detailed analysis on LSTM memory retaining, and a benchmarking of nine different pooling methods on LSTM models using 1.7M generated mixture clips of multiple events with different signal-to-noise ratios. This paper focuses on understanding: 1) utterance-level classification accuracy; 2) sensitivity to event position within an utterance. The analysis is done on the dataset for the detection of rare sound events from DCASE 2017 Challenge. We find max pooling on the prediction level to perform the best among the nine pooling approaches in terms of classification accuracy and insensitivity to event position within an utterance. To authors' best knowledge, this is the first kind of such work focused on LSTM dynamics for AEC tasks.

arXiv:2001.11278

A Rule-Based EEG Classification System for Discrimination of Hand Motor Attempts in Stroke Patients

Stroke patients have symptoms of cerebral functional disturbance that could aggressively impair patient's physical mobility, such as freezing of hand movements. Although rehabilitation training from external devices is beneficial for hand movement recovery, for initiating motor function restoration purposes, there are still valuable research merits for identifying the side of hands in motion. In this preliminary study, we used electroencephalogram (EEG) datasets from 8 stroke patients, with each subject involving 40 EEG trials of left motor attempts and 40 EEG trials of right motor attempts. Then, we proposed a rule-based EEG classification system for identifying the side in motion for stroke patients. In specific, we extracted 1-50 Hz power spectral features as input features of a series of well-known classification models. The predicted labels from these classification models were measured by four types of rules, which determined the finalised predicted label. Our experiment results showed that our proposed rule-based EEG classification system achieved $99.83 \pm 0.42 \% $ accuracy, $ 99.98 \pm 0.13\% $ precision, $ 99.66 \pm 0.84 \% $ recall, and $ 99.83 \pm 0.43\% $ f-score, which outperformed the performance of single well-known classification models. Our findings suggest that the superior performance of our proposed rule-based EEG classification system has the potential for hand rehabilitation in stroke patients.

arXiv:2001.10902

RPCA-Based High Resolution Through-the-Wall Human Motion Detection and Classification

Radar based assisted living has received great amount of research interest in recent years. By employing the micro-Doppler features of indoor human motions, accurate recognition and classification of different types of movements become possible. Whereas, most of the existing works are focused only on free space detection, the literature on detection and recognition of human motions in through-the-wall scenarios is still in its infancy. As can be anticipated, the wall media and indoor static non-human targets would cause clutters and significantly corrupt the motion information of human subjects behind wall. However, no relevant work is reported to effectively handle this problem. In the present work, we aim to fill the gap and propose to use a low center-frequency ultra-wideband (UWB) radar system to probe the behind wall scene. Then, a Robust Principal Component Analysis (RPCA) based subspace decomposition technique, as its first reported implementation, is employed not only to remove the stationary clutters in raw range slow-time map but also to mitigate the multipath effects in the time-frequency map. Onsite experiments of detecting human motions behind a single layer of concrete wall is carried out to investigate the performance of the technique. Lastly, a two dimensional (2D)-PCA algorithm-based motion classification is provided to further verify the effectiveness of the proposed technique. Classification result shows that an enhanced recognition capability can be achieved using the proposed technique in detection and classification of indoor human motions.

arXiv:2001.10449

Range-Max Enhanced Ultra-Wideband Micro-Doppler Signatures of Behind Wall Indoor Human Activities

Penetrating detection and recognition of behind wall indoor human activities has drawn great attentions from social security and emergency service department in recent years since intelligent surveillance aforehand could avail the proper decision making before operations being carried out. However, due to the influence of the wall effects, the obtained micro-Doppler signatures would be severely degenerated by strong near zero-frequency DC components, which would inevitably smear the detailed characteristic features of different behind wall motions in time-frequency (TF) map and further hinder the motion recognition and classification. In this paper, an ultra-wideband (UWB) radar system is first employed to probe through the opaque wall to detect the behind wall motions, which often span a certain number of range bin cells. By employing such a system, a high resolution range map can be obtained, in which the embedded rich range information is expected to be fully exploited to improve the subsequent recognition and classification performance. Secondly, a high-pass filter is applied to remove the effect of the wall in the raw range map. Then, with the aim of enhancing the characteristic features of different behind wall motions in TF maps, a novel range-max enhancement strategy is proposed to extract the most significant micro-Doppler feature of each TF cell along all range bins for a specific motion. Lastly, the effectiveness of the proposed micro-Doppler signature enhancement strategy is investigated by means of onsite experiments and comparative classification. Both the feature enhanced TF maps and classification results show that the proposed approach outperforms other state-of-art Short-Time Fourier Transform (STFT) based TF feature extraction methods.

arXiv:2001.04225공식 코드 · Papers with Code

Evaluation of convolutional neural networks using a large multi-subject P300 dataset

Deep neural networks (DNN) have been studied in various machine learning areas. For example, event-related potential (ERP) signal classification is a highly complex task potentially suitable for DNN as signal-to-noise ratio is low, and underlying spatial and temporal patterns display a large intra- and intersubject variability. Convolutional neural networks (CNN) have been compared with baseline traditional models, i.e. linear discriminant analysis (LDA) and support vector machines (SVM) for single trial classification using a large multi-subject publicly available P300 dataset of school-age children (138 males and 112 females). For single trial classification, classification accuracy stayed between 62% and 64% for all tested classification models. When applying the trained classification models to averaged trials, accuracy increased to 76-79% without significant differences among classification models. CNN did not prove superior to baseline for the tested dataset. Comparison with related literature, limitations and future directions are discussed.

arXiv:1912.10442

End-Point Detection with State Transition Model based on Chunk-Wise Classification

A state transition model (STM) based on chunk-wise classification was proposed for end-point detection (EPD). In general, EPD is developed using frame-wise voice activity detection (VAD) with additional STM, in which the state transition is conducted based on VAD's frame-level decision (speech or non-speech). However, VAD errors frequently occur in noisy environments, even though we use state-of-the-art deep neural network based VAD, which causes the undesired state transition of STM. In this work, to build robust STM, a state transition is conducted based on chunk-wise classification as EPD does not need to be conducted in frame-level. The chunk consists of multiple frames and the classification of chunk between speech and non-speech is done by aggregating the decisions of VAD for multiple frames, so that some undesired VAD errors in a chunk can be smoothed by other correct VAD decisions. Finally, the model was evaluated in both qualitative and quantitative measures including phone error rate.

arXiv:1910.11400

Meta-learning for robust child-adult classification from speech

Computational modeling of naturalistic conversations in clinical applications has seen growing interest in the past decade. An important use-case involves child-adult interactions within the autism diagnosis and intervention domain. In this paper, we address a specific sub-problem of speaker diarization, namely child-adult speaker classification in such dyadic conversations with specified roles. Training a speaker classification system robust to speaker and channel conditions is challenging due to inherent variability in the speech within children and the adult interlocutors. In this work, we propose the use of meta-learning, in particular, prototypical networks which optimize a metric space across multiple tasks. By modeling every child-adult pair in the training set as a separate task during meta-training, we learn a representation with improved generalizability compared to conventional supervised learning. We demonstrate improvements over state-of-the-art speaker embeddings (x-vectors) under two evaluation settings: weakly supervised classification (up to 14.53% relative improvement in F1-scores) and clustering (up to relative 9.66% improvement in cluster purity). Our results show that protonets can potentially extract robust speaker embeddings for child-adult classification from speech.

arXiv:1910.00071

Interpretable Convolutional Neural Networks for Preterm Birth Classification

The use of convolutional neural networks (CNNs) for classification tasks has become dominant in various medical imaging applications. At the same time, recent advances in interpretable machine learning techniques have shown great potential in explaining classifiers' decisions. Layer-wise relevance propagation (LRP) has been introduced as one of these novel methods that aim to provide visual interpretation for the network's decisions. In this work we propose the application of 3D CNNs with LRP for the first time for neonatal T2-weighted magnetic resonance imaging (MRI) data analysis. Through LRP, the decisions of our trained classifier are transformed into heatmaps indicating each voxel's relevance for the outcome of the decision. Our resulting LRP heatmaps reveal anatomically plausible features in distinguishing preterm neonates from term ones.