Recent research on model interpretability in natural language processing extensively uses feature scoring methods for identifying which parts of the input are the most important for a model to make a prediction (i.e. explanation or rationale). However, previous research has shown that there is no clear best scoring method across various text classification tasks while practitioners typically have to make several other ad-hoc choices regarding the length and the type of the rationale (e.g. short or long, contiguous or not). Inspired by this, we propose a simple yet effective and flexible method that allows selecting optimally for each data instance: (1) a feature scoring method; (2) the length; and (3) the type of the rationale. Our method is inspired by input erasure approaches to interpretability which assume that the most faithful rationale for a prediction should be the one with the highest difference between the model's output distribution using the full text and the text after removing the rationale as input respectively. Evaluation on four standard text classification datasets shows that our proposed method provides more faithful, comprehensive and highly sufficient explanations compared to using a fixed feature scoring method, rationale length and type. More importantly, we demonstrate that a practitioner is not required to make any ad-hoc choices in order to extract faithful rationales using our approach.
A health outcome is a measurement or an observation used to capture and assess the effect of a treatment. Automatic detection of health outcomes from text would undoubtedly speed up access to evidence necessary in healthcare decision making. Prior work on outcome detection has modelled this task as either (a) a sequence labelling task, where the goal is to detect which text spans describe health outcomes, or (b) a classification task, where the goal is to classify a text into a pre-defined set of categories depending on an outcome that is mentioned somewhere in that text. However, this decoupling of span detection and classification is problematic from a modelling perspective and ignores global structural correspondences between sentence-level and word-level information present in a given text. To address this, we propose a method that uses both word-level and sentence-level information to simultaneously perform outcome span detection and outcome type classification. In addition to injecting contextual information to hidden vectors, we use label attention to appropriately weight both word and sentence level information. Experimental results on several benchmark datasets for health outcome detection show that our proposed method consistently outperforms decoupled methods, reporting competitive results.
One of the challenges for current sequence to sequence (seq2seq) models is processing long sequences, such as those in summarization and document level machine translation tasks. These tasks require the model to reason at the token level as well as the sentence and paragraph level. We design and study a new Hierarchical Attention Transformer-based architecture (HAT) that outperforms standard Transformers on several sequence to sequence tasks. Furthermore, our model achieves state-of-the-art ROUGE scores on four summarization tasks, including PubMed, arXiv, CNN/DM, SAMSum, and AMI. Our model outperforms document-level machine translation baseline on the WMT20 English to German translation task. We investigate what the hierarchical layers learn by visualizing the hierarchical encoder-decoder attention. Finally, we study hierarchical learning on encoder-only pre-training and analyze its performance on classification tasks.
This paper describes our submissions for the Social Media Mining for Health (SMM4H)2021 shared tasks. We participated in 2 tasks:(1) Classification, extraction and normalization of adverse drug effect (ADE) mentions in English tweets (Task-1) and (2) Classification of COVID-19 tweets containing symptoms(Task-6). Our approach for the first task uses the language representation model RoBERTa with a binary classification head. For the second task, we use BERTweet, based on RoBERTa. Fine-tuning is performed on the pre-trained models for both tasks. The models are placed on top of a custom domain-specific processing pipeline. Our system ranked first among all the submissions for subtask-1(a) with an F1-score of 61%. For subtask-1(b), our system obtained an F1-score of 50% with improvements up to +8% F1 over the score averaged across all submissions. The BERTweet model achieved an F1 score of 94% on SMM4H 2021 Task-6.
The Hierarchical Vote Collective of Transformation-based Ensembles (HIVE-COTE) is a heterogeneous meta ensemble for time series classification. HIVE-COTE forms its ensemble from classifiers of multiple domains, including phase-independent shapelets, bag-of-words based dictionaries and phase-dependent intervals. Since it was first proposed in 2016, the algorithm has remained state of the art for accuracy on the UCR time series classification archive. Over time it has been incrementally updated, culminating in its current state, HIVE-COTE 1.0. During this time a number of algorithms have been proposed which match the accuracy of HIVE-COTE. We propose comprehensive changes to the HIVE-COTE algorithm which significantly improve its accuracy and usability, presenting this upgrade as HIVE-COTE 2.0. We introduce two novel classifiers, the Temporal Dictionary Ensemble (TDE) and Diverse Representation Canonical Interval Forest (DrCIF), which replace existing ensemble members. Additionally, we introduce the Arsenal, an ensemble of ROCKET classifiers as a new HIVE-COTE 2.0 constituent. We demonstrate that HIVE-COTE 2.0 is significantly more accurate than the current state of the art on 112 univariate UCR archive datasets and 26 multivariate UEA archive datasets.
Audio classification using breath and cough samples has recently emerged as a low-cost, non-invasive, and accessible COVID-19 screening method. However, a comprehensive survey shows that no application has been approved for official use at the time of writing, due to the stringent reliability and accuracy requirements of the critical healthcare setting. To support the development of Machine Learning classification models, we performed an extensive comparative investigation and ranking of 15 audio features, including less well-known ones. The results were verified on two independent COVID-19 sound datasets. By using the identified top-performing features, we have increased COVID-19 classification accuracy by up to 17% on the Cambridge dataset and up to 10% on the Coswara dataset compared to the original baseline accuracies without our feature ranking.
We examine the use of linear and non-linear dimensionality reduction algorithms for extracting low-rank feature representations for speech emotion recognition. Two feature sets are used, one based on low-level descriptors and their aggregations (IS10) and one modeling recurrence dynamics of speech (RQA), as well as their fusion. We report speech emotion recognition (SER) results for learned representations on two databases using different classification methods. Classification with low-dimensional representations yields performance improvement in a variety of settings. This indicates that dimensionality reduction is an effective way to combat the curse of dimensionality for SER. Visualization of features in two dimensions provides insight into discriminatory abilities of reduced feature sets.
The performance of text classification methods has improved greatly over the last decade for text instances of less than 512 tokens. This limit has been adopted by most state-of-the-research transformer models due to the high computational cost of analyzing longer text instances. To mitigate this problem and to improve classification for longer texts, researchers have sought to resolve the underlying causes of the computational cost and have proposed optimizations for the attention mechanism, which is the key element of every transformer model. In our study, we are not pursuing the ultimate goal of long text classification, i.e., the ability to analyze entire text instances at one time while preserving high performance at a reasonable computational cost. Instead, we propose a text truncation method called Text Guide, in which the original text length is reduced to a predefined limit in a manner that improves performance over naive and semi-naive approaches while preserving low computational costs. Text Guide benefits from the concept of feature importance, a notion from the explainable artificial intelligence domain. We demonstrate that Text Guide can be used to improve the performance of recent language models specifically designed for long text classification, such as Longformer. Moreover, we discovered that parameter optimization is the key to Text Guide performance and must be conducted before the method is deployed. Future experiments may reveal additional benefits provided by this new method.
Sleep disorder is one of many neurological diseases that can affect greatly the quality of daily life. It is very burdensome to manually classify the sleep stages to detect sleep disorders. Therefore, the automatic sleep stage classification techniques are needed. However, the previous automatic sleep scoring methods using raw signals are still low classification performance. In this study, we proposed an end-to-end automatic sleep staging framework based on optimal spectral-temporal sleep features using a sleep-edf dataset. The input data were modified using a bandpass filter and then applied to a convolutional neural network model. For five sleep stage classification, the classification performance 85.6% and 91.1% using the raw input data and the proposed input, respectively. This result also shows the highest performance compared to conventional studies using the same dataset. The proposed framework has shown high performance by using optimal features associated with each sleep stage, which may help to find new features in the automatic sleep stage method.
Multitarget tracking (MTT) is a challenging task that aims at estimating the number of targets and their states from measurements of the target states provided by one or multiple sensors. Additional information, such as imperfect estimates of target classes provided by a classifier, can facilitate the target-measurement association and thus improve MTT performance. In this letter, we describe how a recently proposed MTT framework based on the sum-product algorithm can be extended to efficiently exploit class information. The effectiveness of the proposed approach is demonstrated by simulation results.
This paper addresses an autonomous exploration problem in which a mobile sensor, placed in a previously unseen search area, utilizes an information-theoretic navigation cost function to dynamically select the next sensing action, i.e., location from which to take a measurement, to efficiently detect and classify objects of interest within the area. The information-theoretic cost function proposed in this paper consist of two \textit{information gain} terms, one for detection and localization of objects and the other for sequential classification of the detected objects. We present a novel closed-form representation for the cost function, derived from the definition of mutual information. We evaluate three different policies for choosing the next sensing action: lawn mower, greedy, and non-greedy. For these three policies, we compare the results from our information-theoretic cost functions to the results of other information-theoretic inspired cost functions. Our simulation results show that search efficiency is greater using the proposed cost functions compared to those of the other methods, and that the greedy and non-greedy policies outperform the lawn mower policy.
Electrocardiogram (ECG) is the most crucial monitoring modality to diagnose cardiovascular events. Precise and automatic detection of abnormal ECG patterns is beneficial to both physicians and patients. In the automatic detection of abnormal ECG patterns, deep neural networks (DNNs) have shown significant achievements. However, DNNs require large amount of labeled data, which are often expensive to obtain. On the other hand, recent research have shown by randomly combining data augmentations can improve image classification accuracy. Thus, in this work we explore data augmentation suitable for ECG data and propose ECG Augment. We show by introducing ECG Augment, we can improve classification of atrial fibrillation with single lead ECG data, without changing an architecture of DNN.
Modulation classification, recognized as the intermediate step between signal detection and demodulation, is widely deployed in several modern wireless communication systems. Although many approaches have been studied in the last decades for identifying the modulation format of an incoming signal, they often reveal the obstacle of learning radio characteristics for most traditional machine learning algorithms. To overcome this drawback, we propose an accurate modulation classification method by exploiting deep learning for being compatible with constellation diagram. Particularly, a convolutional neural network is developed for proficiently learning the most relevant radio characteristics of gray-scale constellation image. The deep network is specified by multiple processing blocks, where several grouped and asymmetric convolutional layers in each block are organized by a flow-in-flow structure for feature enrichment. These blocks are connected via skip-connection to prevent the vanishing gradient problem while effectively preserving the information identify throughout the network. Regarding several intensive simulations on the constellation image dataset of eight digital modulations, the proposed deep network achieves the remarkable classification accuracy of approximately 87% at 0 dB signal-to-noise ratio (SNR) under a multipath Rayleigh fading channel and further outperforms some state-of-the-art deep models of constellation-based modulation classification.
Modulation classification, an intermediate process between signal detection and demodulation in a physical layer, is now attracting more interest to the cognitive radio field, wherein the performance is powered by artificial intelligence algorithms. However, most existing conventional approaches pose the obstacle of effectively learning weakly discriminative modulation patterns. This paper proposes a robust modulation classification method by taking advantage of deep learning to capture the meaningful information of modulation signal at multi-scale feature representations. To this end, a novel architecture of convolutional neural network, namely Chain-Net, is developed with various asymmetric kernels organized in two processing flows and associated via depth-wise concatenation and element-wise addition for optimizing feature utilization. The network is evaluated on a big dataset of 14 challenging modulation formats, including analog and high-order digital techniques. The simulation results demonstrate that Chain-Net robustly classifies the modulation of radio signals suffering from a synthetic channel deterioration and further performs better than other deep networks.
In this paper we are interested to learn from a wireless digitally modulated signal the number of antennas that the transmitter (Tx) of this signal uses, as well as its specific modulation scheme (from phase-shift keying (PSK) or quadrature amplitude modulation (QAM)). Formally, these are modulation and antenna classification problems. We examine the problems with data-driven machine learning (ML)-based techniques. The two sub-problems of modulation and number of transmitter antenna classification are initially examined independently for a variety for system parameters, namely the SNR, number of receiver (Rx) antennas, and classification algorithms. Then we consider the joint problem where we follow two approaches. One, where the sub-problems are solved independently and in parallel, and one where the antenna classifier waits on the result of the modulation classifier. The two proposed schemes do not require any knowledge/details of the used modulation schemes and the way the Tx antennas are used (spatial multiplexing, space-time codes,etc.) as it is fully data-driven and not decision-theoretic based. The results of our approach are characterized by high classification accuracy and they pave the way for more ML-based data-driven techniques that reveal more characteristics of the Tx.
Mental task identification and classification using single/limited channel(s) electroencephalogram (EEG) signals in real-time play an important role in the design of portable brain-computer interface (BCI) and neurofeedback (NFB) systems. However, the real-time recorded EEG signals are often contaminated with noises such as ocular artifacts (OAs) and muscle artifacts (MAs), which deteriorate the hand-crafted features extracted from EEG signal, resulting inadequate identification and classification of mental tasks. Therefore, we investigate the use of recent deep learning techniques which do not require any manual feature extraction or artifact suppression step. In this paper, we propose a light-weight one-dimensional convolutional neural network (1D-CNN) architecture for mental task identification and classification. The robustness of the proposed architecture is evaluated using artifact-free and artifact-contaminated EEG signals taken from two publicly available databases (i.e, Keirn and Aunon ($K$) database and EEGMAT ($E$) database) and in-house ($R$) database recorded using single-channel neurosky mindwave mobile 2 (MWM2) EEG headset in performing not only mental/non-mental binary task classification but also different mental/mental multi-tasks classification. Evaluation results demonstrate that the proposed architecture achieves the highest subject-independent classification accuracy of $99.7\%$ and $100\%$ for multi-class classification and pair-wise mental tasks classification respectively in database $K$. Further, the proposed architecture achieves subject-independent classification accuracy of $99\%$ and $98\%$ in database $E$ and the recorded database $R$ respectively. Comparative performance analysis demonstrates that the proposed architecture outperforms existing approaches not only in terms of classification accuracy but also in robustness against artifacts.
Detecting UAVs is becoming more crucial for various industries such as airports and nuclear power plants for improving surveillance and security measures. Exploiting radio frequency (RF) based drone control and communication enables a passive way of drone detection for a wide range of environments and even without favourable line of sight (LOS) conditions. In this paper, we evaluate RF based drone classification performance of various state-of-the-art (SoA) models on a new realistic drone RF dataset. With the help of a newly proposed residual Convolutional Neural Network (CNN) model, we show that the drone RF frequency signatures can be used for effective classification. The robustness of the classifier is evaluated in a multipath environment considering varying Doppler frequencies that may be introduced from a flying drone. We also show that the model achieves better generalization capabilities under different wireless channel and drone speed scenarios. Furthermore, the newly proposed model's classification performance is evaluated on a simultaneous multi-drone scenario. The classifier achieves close to 99 % classification accuracy for signal-to-noise ratio (SNR) 0 dB and at -10 dB SNR it obtains 5 % better classification accuracy compared to the existing framework.
Brain-Computer Interfaces (BCI) based on motor imagery translate mental motor images recognized from the electroencephalogram (EEG) to control commands. EEG patterns of different imagination tasks, e.g. hand and foot movements, are effectively classified with machine learning techniques using band power features. Recently, also Convolutional Neural Networks (CNNs) that learn both effective features and classifiers simultaneously from raw EEG data have been applied. However, CNNs have two major drawbacks: (i) they have a very large number of parameters, which thus requires a very large number of training examples; and (ii) they are not designed to explicitly learn features in the frequency domain. To overcome these limitations, in this work we introduce Sinc-EEGNet, a lightweight CNN architecture that combines learnable band-pass and depthwise convolutional filters. Experimental results obtained on the publicly available BCI Competition IV Dataset 2a show that our approach outperforms reference methods in terms of classification accuracy.
While RFML is expected to be a key enabler of future wireless standards, a significant challenge to the widespread adoption of RFML techniques is the lack of explainability in deep learning models. This work investigates the use of CB models as a means to provide inherent decision explanations in the context of DL-based AMC. Results show that the proposed approach not only meets the performance of single-network DL-based AMC algorithms, but provides the desired model explainability and shows potential for classifying modulation schemes not seen during training (i.e. zero-shot learning).
Classification of cognitive workload promises immense benefit in diverse areas ranging from driver safety to augmenting human capability through closed loop brain computer interface. The brain is the most metabolically active organ in the body and increases its metabolic activity and thus oxygen consumption with increasing cognitive demand. In this study, we explore the feasibility of in-ear SpO2 cognitive workload tracking. To this end, we preform cognitive workload assessment in 8 subjects, based on an N-back task, whereby the subjects are asked to count and remember the number of odd numbers displayed on a screen in 5 second windows. The 2 and 3-back tasks lead to either the lowest median absolute SpO2 or largest median decrease in SpO2 in all of the subjects, indicating a robust and measurable decrease in blood oxygen in response to increased cognitive workload. Using features derived from in-ear pulse oximetry, including SpO2, pulse rate and respiration rate, we were able to classify the 4 N-back task categories, over 5 second epochs, with a mean accuracy of 94.2%. Moreover, out of 21 total features, the 9 most important features for classification accuracy were all SpO2 related features. The findings suggest that in-ear SpO2 measurements provide valuable information for classification of cognitive workload over short time windows, which together with the small form factor promises a new avenue for real time cognitive workload tracking.