KNN Classifier Research Articles

Detection of lung cancer using novel attention gate residual U-Net model and KNN classifier from computer tomography images

Computed tomography (CT) scan pictures are routinely employed in the automatic identification and classification of lung cancer. The texture distribution of lung nodules can vary widely over the CT scan space and requires accurate detection. The evaluation of discriminative information in this volume can tremendously aid the classification process. A convolutional neural network, the Attention Gate Residual U-Net model, and KNN classifiers are utilized to detect lung cancer. The dataset of 1097 computed tomography (CT) images utilized in this study was obtained from the Iraq-Oncology Teaching Hospital/National Centre for Cancer Diseases (IQ-OTH/NCCD) to segment and classify lung tumors from CT images using the novel Attention Gate Residual U-Net model, i.e., AGResU-Net and CNN architecture. The initial step is applying CNN to detect normal, benign, and malignant patients in CT images. Second, use AGResU-Net to partition lung tumour areas. In the third section of the project, a KNN classifier is used to determine if an instance is malignant or benign. In the initial phase, CNN was proposed to classify three distinct regions. Three optimization strategies are used in this work: Adam, RMSP, and SGDM. The classifier’s accuracy is 97%, 85%, and 82%, respectively. When compared to the RMSP optimizer, the Adams optimizer predicts probability rates more accurately. In the second phase, AGResU-Net is used for schematic segmentation of the tumor region. In the third phase, a KNN classifier is used to classify benign and malignant tumor from the segmented tumor regions. A new segmentation of the lung tumor model is proposed. In this developed algorithm, the labelled classified data set and the segmented tumor output result provide the same accuracy. The study results demonstrate high tumour classification accuracy and high probability of detection in benign and malignant cases.

Classification of ischemic and hemorrhagic stroke using Enhanced-CNN deep learning technique

Stroke is a type of cerebrovascular disorder that has a significant impact on people’s lives and well-being. Quantitative investigation of MRI imaging of the brain plays a critical role in analyzing and identifying therapy for stroke. A block primarily provokes stroke in the brain’s blood supply. Deep learning algorithms can be used to identify strokes in patients in a short period. Proposed deep learning methods are used to classify strokes using magnetic resonance imaging (MRI) images. Early detection enhances treatment opportunities and saves lives, which is the primary motivation of the proposed work. Deep learning methods have emerged as significant research trends in recent years, particularly for classifying different types of stroke such as ischemic and hemorrhagic stroke. A dataset of 13,850 MRI images of stroke patients was collected from various reliable sources, including Madras scans and labs, Radiopaedia, Kaggle datasets, and online databases. Among these images, 7,810 were identified as cases of ischemic stroke, while 6,040 represented hemorrhagic strokes. For training purposes, a total of 9,700 images were used, with 4,150 images employed for testing. A comparative analysis of ANN, SVM, NB, ELM, KNN and Enhanced CNN technique is carried out, and 98.4% of classification accuracy is obtained by using Enhanced CNN. Statistical analysis of parameters such as accuracy, precision, F1-score, and recall was conducted, demonstrating that the Enhanced CNN method outperformed SVM, NB,ELM, KNN and ANN classifiers. The Enhanced CNN method achieved an accuracy of 0.984, precision of 0.949, recall of 0.972, and an F1-score of 0.960 on the training dataset, which is significantly higher than the other classifiers. Furthermore, the Enhanced CNN algorithm’s ability to automatically learn features and efficiently process large datasets enhances its potential as a powerful tool for accurately classifying stroke lesions.

Realtime Evaluation of Cardiac Activities Using Multi Features with KNN Classifier - A Pilot Study

Monitoring of cardiac episodes is essential for patients with suspected cardiac arrhythmias. The quantitative evaluation of cardiac activities during continuous unattended cardiac monitoring will support clinician to understand the nature of the physical activities. This pilot study is proposed to assess the cardiac activities using multi features with KNN classifier. ECG was collected from healthy volunteers under rest and during exercise. The time and frequency domain features such as BPM, QRS duration, RR interval, SD, CV, RMSRR, NN50, SDNN, Shannon entropy, log entropy, and sample entropy were extracted. Further, multi-classification was employed using KNN classifier. The performance of KNN classifier is tested in terms of classification accuracy and elapsed time and it was found that Euclidean distance for k=2 was showing better performance. Statistical analysis was performed to confirm the suitability of the features using Pearson’s correlation coefficient and box plot. Among the features BPM, QRS, RR, sample and log entropies were found to be significant. Specifically, sample entropy provided to be the best candidate with p<0.01. The study reveals an overall classification accuracy of 90% was obtained.

Development of Real-Time Emotion Recognition System Using Facial Expressions

This research presents a real-time emotion recognition system that combines human-friendly machine interaction with picture processing. For many years, facial detection has been available. Moving further, it is possible to simulate the emotions that people express on their faces and experience in their brains through the use of video, electric signals, or image forms. Since it is hard for computers to detect emotions from images or videos and a difficult task for the human eye, machine emotion detection requires a variety of image processing approaches for feature extraction. The approach proposed in this paper consists of two primary processes: facial expression recognition (FER) and face detection. The experimental investigation of facial emotion recognition is the main topic of this study. An emotion detection system's workflow consists of face detection, feature extraction, pre-processing, classification, and image acquisition. The emotion identification system uses the Haar cascade algorithm, an object detection algorithm, to recognize faces in an image or a real-time video, and the KNN Classifier for image classification in order to identify such emotions. Using the webcam to capture real-time photos, this system operates. The goal of this research is to develop an automatic facial expression recognition system that can recognize various emotions. Based on these studies, the system may be able to distinguish between a number of people who are fearful, furious, shocked, sad, or pleased, among other emotions.

Comparison of Sentiment Analysis Methods on Topic Haram of Music In Youtube

Sentiment analysis on video lectures on YouTube that discuss the haram of music is an exciting topic to find out public opinion. This study aims to find what factors affect the model's accuracy in sentiment analysis, especially on video lecture content on YouTube. The data used is comment data on three video lectures that discuss the haram of music, which has been labelled into two categories: positive and negative. The data is divided into two categories, namely primary data, as many as 2099 data that have not been normalized, while secondary data has 1001 data that have been normalized. The experiment shows that the validity of the data, labelling the data, the amount of data, and preprocessing are essential points in forming a good sentiment analysis classification model because, from the test results, it was found that imbalance techniques such as SMOTE, word embedding word2Vec and FastText, and SVM and KNN classification algorithms do not provide maximum accuracy if the data used primary data. However, the data imbalance process, such as oversampling and SVM and KNN classification algorithms, will provide better model accuracy if used with secondary data. Based on the trial results, it is found that when using the SVM algorithm, primary data produces the highest accuracy at 58.35%, while secondary data is 72.23%. If using KNN, the primary data provides the highest model accuracy at 53.54%, while the secondary data has the highest accuracy at 72.81%. Based on these results, it was found that the validity of the data or data must be appropriate and related to the case raised and labelling the data must be done carefully because the most crucial is the inappropriate data in preprocessing the data must be done correctly, if data preprocessing is done in an inappropriate way then data imbalance techniques such as oversampling do not have enough influence on increasing accuracy, but if on the contrary then accuracy will increase. The selection of the right word embedding also affects accuracy. It is necessary to do many experiments to select the correct algorithm and follow the data owned because selecting the correct algorithm will provide maximum accuracy model results

Feature selection using a sinusoidal sequence combined with mutual information

Data classification is the most common task in machine learning, and feature selection is the key step in the classification task. Common feature selection methods mainly analyze the maximum correlation and minimum redundancy between feature factors and tags while ignoring the impact of the number of key features, which will inevitably lead to waste in subsequent classification training. To solve this problem, a feature selection algorithm (SSMI) based on the combination of sinusoidal sequences and mutual information is proposed. First, the mutual information between each feature and tag is calculated, and the interference information in high-dimensional data is removed according to the mutual information value. Second, a sine function is constructed, and sine ordering is carried out according to the mutual information value and feature mean value between different categories of the same feature. By adjusting the period and phase value of the sequence, the feature set with the largest difference is found, and the subset of key features is obtained. Finally, three machine learning classifiers (KNN, RF, SVM) are used to classify key feature subsets, and several feature selection algorithms (JMI, mRMR, CMIM, SFS, etc.) are compared to verify the advantages and disadvantages of different algorithms. Compared with other feature selection methods, the SSMI algorithm obtains the least number of key features, with an average reduction of 15 features. The average classification accuracy has been improved by 3% on the KNN classifier. On the HBV and SDHR datasets, the SSMI algorithm achieved classification accuracy of 81.26% and 83.12%, with sensitivity and specificity results of 76.28%, 87.39% and 68.14%, 86.11%, respectively. This shows that the SSMI algorithm can achieve higher classification accuracy with a smaller feature subset.

The effect of time correlation function, Sauter mean diameter and asymmetry ratio on machine learning classifiers for particle discrimination by using light scattering

This study illustrates that different particle types can be accurately discriminated using a novel and simple detector cell. ML models for the classification purpose were trained with scattered light at different wavelengths. Different smokes of wood, polyurethane, sunflower oil and paraffin, PAO, DEHS aerosols were introduced to the measurement cell. Measurements for each material were repeated ten times at five different test scenarios. The measurement range is up to 1.42% obs/m. For particle analysis, asymmetry ratio, Sauter mean diameter information and TCF were derived from the scattered light. NN, RFC, SVC and KNN classifiers were trained with different features to investigate the effect of each feature and a comparison study on the classification performances of ML models was carried out. ML models which include all features, identified smoke of wood and polyurethane with high accuracies. According to the F1 scores, the top two classifiers were determined as NN and SVC. The F1 scores of the proposed best models ranged from 71.0% to 100.0%. TCF data generally improve the classification performance. Proposed best ML models are capable of discriminating smoke types at the early stage of fire and can be tailored to new particle types and provide excellent quantitative accuracy.

Stratifying knee osteoarthritis features through multitask deep hybrid learning: Data from the osteoarthritis initiative

Background and objectiveKnee osteoarthritis (OA) is a debilitating musculoskeletal disorder that causes functional disability. Automatic knee OA diagnosis has great potential of enabling timely and early intervention, that can potentially reverse the degenerative process of knee OA. Yet, it is a tedious task, concerning the heterogeneity of the disorder. Most of the proposed techniques demonstrated single OA diagnostic task widely based on Kellgren Lawrence (KL) standard, a composite score of only a few imaging features (i.e. osteophytes, joint space narrowing and subchondral bone changes). However, only one key disease pattern was tackled. The KL standard fails to represent disease pattern of individual OA features, particularly osteophytes, joint-space narrowing, and pain intensity that play a fundamental role in OA manifestation. In this study, we aim to develop a multitask model using convolutional neural network (CNN) feature extractors and machine learning classifiers to detect nine important OA features: KL grade, knee osteophytes (both knee, medial fibular: OSFM, medial tibial: OSTM, lateral fibular: OSFL, and lateral tibial: OSTL), joint-space narrowing (medial: JSM, and lateral: JSL), and patient-reported pain intensity from plain radiography. MethodsWe proposed a new feature extraction method by replacing fully-connected layer with global average pooling (GAP) layer. A comparative analysis was conducted to compare the efficacy of 16 different convolutional neural network (CNN) feature extractors and three machine learning classifiers. ResultsExperimental results revealed the potential of CNN feature extractors in conducting multitask diagnosis. Optimal model consisted of VGG16-GAP feature extractor and KNN classifier. This model not only outperformed the other tested models, it also outperformed the state-of-art methods with higher balanced accuracy, higher Cohen's kappa, higher F1, and lower mean squared error (MSE) in seven OA features prediction. ConclusionsThe proposed model demonstrates pain prediction on plain radiographs, as well as eight OA-related bony features. Future work should focus on exploring additional potential radiological manifestations of OA and their relation to therapeutic interventions.

An integrated approach ofIoTandWSNusing wavelet transform and machine learning for the solid waste image classification in smart cities

AbstractSolid waste management is one of the utmost vital issues allied with the smart city, which harms our society's health and the environment. Solid waste management has been a significant hurdle in developing smart cities, impacting the community's living standards. Sustainability is the circumstances that help to achieve natural and social development simultaneously in a prolific accord. We can allege that sustainable growth brings out stability in the necessities of the environment. It helps to preserve resources available for the coming generations. Waste management is also a concern for many authorities; researchers have started focusing on this area. This research article presents a unique method for waste management by employing internet of things for waste monitoring, a wireless sensor network for data communication and discrete wavelet transforms, image processing, and machine learning for image categorization. Implementing machine learning for image categorization has gained attention in the research area of artificial intelligence. The paper presents a way to classify waste images into different Classes such as domestic waste, e‐waste, and medical waste. The input image is down‐sampled, and channels are split to extract the color moments for minimal processing time. We have created a convolution neural network (CNN) trained using pre‐processed data for feature extraction and classification. We have used K‐nearest neighbor classifier (KNN) as another classifier to compare the results obtained by the proposed methodology. The findings demonstrate that, compared to the KNN, the suggested model shortens the processing time and achieves a high accuracy rate. The paper presents a novel approach for Solid waste classification by using pre‐processed image data for image processing and CNN for waste image classification, for the sake of proving the efficiency of the proposed method we have compared the results with the KNN classifier and found that the accuracy of classification is better in proposed method than the KNN classifier.

Semi-supervised labelling of chest x-ray images using unsupervised clustering for ground-truth generation

Supervised classifiers require a lot of data with accurate labels to learn to recognize chest X-ray images (CXR). However, manually labeling an extensive collection of CXR images is time-consuming and costly. To address this issue, a method for the semi-supervised labelling of extensive collections of CXR images is proposed leveraging unsupervised clustering with minimum expert knowledge to generate ground truth images. The proposed methodology entails: using unsupervised clustering techniques such as K-Means and Self-Organizing Maps. Second, the images are fed to five different feature vectors to utilize the potential differences between features to their full advantage. Third, each data point gets the label of the cluster’s center to which it belongs. Finally, a majority vote is used to decide the ground truth image. The number of clusters created by the method chosen strictly limits the amount of human involvement. To evaluate the effectiveness of the proposed method, experiments were conducted on two publicly available CXR datasets, namely VinDR-CXR and Montgomery datasets. The experiments showed that, for a KNN classifier, manually labeling only 1% (VinDr-CXR), or 10% (Montgomery) of the training data, gives a similar performance as labeling the whole dataset. The proposed methodology efficiently generates ground-truth images from publicly available CXR datasets. To our knowledge, this is the first study to use the VinDr-CXR and Montgomery datasets for ground truth image generation. Extensive experimental analysis using machine learning and statistical techniques shows that the proposed methodology efficiently generates ground truth images from CXR datasets.

KNN Classification: a review

kNN Classification: a review

ASNET: A Novel AI Framework for Accurate Ankylosing Spondylitis Diagnosis from MRI.

Ankylosing spondylitis (AS) is a chronic, painful, progressive disease usually seen in the spine. Traditional diagnostic methods have limitations in detecting the early stages of AS. The early diagnosis of AS can improve patients' quality of life. This study aims to diagnose AS with a pre-trained hybrid model using magnetic resonance imaging (MRI). In this research, we collected a new MRI dataset comprising three cases. Furthermore, we introduced a novel deep feature engineering model. Within this model, we utilized three renowned pretrained convolutional neural networks (CNNs): DenseNet201, ResNet50, and ShuffleNet. Through these pretrained CNNs, deep features were generated using the transfer learning approach. For each pretrained network, two feature vectors were generated from an MRI. Three feature selectors were employed during the feature selection phase, amplifying the number of features from 6 to 18 (calculated as 6 × 3). The k-nearest neighbors (kNN) classifier was utilized in the classification phase to determine classification results. During the information phase, the iterative majority voting (IMV) algorithm was applied to secure voted results, and our model selected the output with the highest classification accuracy. In this manner, we have introduced a self-organized deep feature engineering model. We have applied the presented model to the collected dataset. The proposed method yielded 99.80%, 99.60%, 100%, and 99.80% results for accuracy, recall, precision, and F1-score for the collected axial images dataset. The collected coronal image dataset yielded 99.45%, 99.20%, 99.70%, and 99.45% results for accuracy, recall, precision, and F1-score, respectively. As for contrast-enhanced images, accuracy of 95.62%, recall of 80.72%, precision of 94.24%, and an F1-score of 86.96% were attained. Based on the results, the proposed method for classifying AS disease has demonstrated successful outcomes using MRI. The model has been tested on three cases, and its consistently high classification performance across all cases underscores the model's general robustness. Furthermore, the ability to diagnose AS disease using only axial images, without the need for contrast-enhanced MRI, represents a significant advancement in both healthcare and economic terms.

Machine learning analysis of lung squamous cell carcinoma gene expression datasets reveals novel prognostic signatures

BackgroundLung squamous cell carcinoma (LUSC) patients are often diagnosed at an advanced stage and have poor prognoses. Thus, identifying novel biomarkers for the LUSC is of utmost importance. MethodsMultiple datasets from the NCBI-GEO repository were obtained and merged to construct the complete dataset. We also constructed a subset from this complete dataset with only known cancer driver genes. Further, machine learning classifiers were employed to obtain the best features from both datasets. Simultaneously, we perform differential gene expression analysis. Furthermore, survival and enrichment analyses were performed. ResultsThe kNN classifier performed comparatively better on the complete and driver datasets' top 40 and 50 gene features, respectively. Out of these 90 gene features, 35 were found to be differentially regulated. Lasso-penalized Cox regression further reduced the number of genes to eight. The median risk score of these eight genes significantly stratified the patients, and low-risk patients have significantly better overall survival. We validated the robust performance of these eight genes on the TCGA dataset. Pathway enrichment analysis identified that these genes are associated with cell cycle, cell proliferation, and migration. ConclusionThis study demonstrates that an integrated approach involving machine learning and system biology may effectively identify novel biomarkers for LUSC.

Highly selective molecularly imprinted polymer nanoparticles (MIP NPs)-based microfluidic gas sensor for tetrahydrocannabinol (THC) detection

A highly selective microfluidic integrated metal oxide gas sensor for THC detection is reported based on MIP nanoparticles (MIP NPs). We synthesized MIP NPs with THC recognition sites and coated them on a 3D-printed microfluidic channel surface. The sensitivity and selectivity of coated microfluidic integrated gas sensors were evaluated by exposure to THC, cannabidiol (CBD), methanol, and ethanol analytes in 300–700 ppm at 300 °C. For comparison, reference signals were obtained from a microfluidic channel coated with nonimprinted polymers (NIP NPs). The MIP and NIP NPs were characterized using scanning electron microscopy (SEM) and Raman spectroscopy. MIP and NIP NPs channels response data were combined and classified with 96.3% accuracy using the Fine KNN classification model in MATLAB R2021b Classification Learner App. Compared to the MIP NPs coated channel, the NIP NPs channel had poor selectivity towards THC, demonstrating that the THC recognition sites in the MIP structure enabled selective detection of THC. The findings demonstrated that the recognition sites of MIP NPs properly captured THC molecules, enabling the selective detection of THC compared to CBD, methanol, and ethanol.

Evolution in pulsating variable stars: Long term, and inter-cycle

Abstract We used machine learning techniques to predict period changes in variable stars using noisy and sparse time series data, while inferring underlying physics and generalizing predictions about cycle-to-cycle variations. Our focus was on Mira variables, a well-known class of pulsating stars. Pre-processing data from Mira, R Andromedae, U Orionis, and Chi Cygni, obtained from the American Association of Variable Star Observers (AAVSO), we predicted luminosity magnitude uncertainty and classified pulsation states. Employing various classification and regression algorithms, along with feature engineering, we aimed to generalize predictions. We created a generalized dataset with collective averaged data points, limiting our analysis to a common time duration. Linear regression models yielded no successful predictions, but Decision Tree and KNN regressors accurately predicted luminosity magnitude errors, indicative of variation over time. Feature engineering successfully aided regression and classification of pulsating star states. After hyper parameter tuning using Bayesian Neural networks, we achieved a classification accuracy of 0.8 and 0.94 for the KNN classifier, respectively, in classifying pulsation states of Mira variables. The regression model achieved an R2 score of 0.98. Our work provides a foundation for developing tools to analyze various pulsating star variables, including Cepheids, RR Lyrae, and Delta Scuti variables, as well as other astrophysical data. These techniques demonstrate impressive performance with time series datasets.

Radiomics and dosiomics for predicting complete response to definitive chemoradiotherapy patients with oesophageal squamous cell cancer using the hybrid institution model.

To develop a multi-institutional prediction model to estimate the local response to oesophageal squamous cell carcinoma (ESCC) treated with definitive radiotherapy based on radiomics and dosiomics features. The local responses were categorised into two groups (incomplete and complete). An external validation model and a hybrid model that the patients from two institutions were mixed randomly were proposed. The ESCC patients at stages I-IV who underwent chemoradiotherapy from 2012 to 2017 and had follow-up duration of more than 5years were included. The patients who received palliative or pre-operable radiotherapy and had no FDG PET images were excluded. The segmentations included the GTV, CTV, and PTV which are used in treatment planning. In addition, shrinkage, expansion, and shell regions were created. Radiomic and dosiomic features were extracted from CT, FDG PET images, and dose distribution. Machine learning-based prediction models were developed using decision tree, support vector machine, k-nearest neighbour (kNN) algorithm, and neural network (NN) classifiers. A total of 116 and 26 patients enrolled at Centre 1 and Centre 2, respectively. The external validation model exhibited the highest accuracy with 65.4% for CT-based radiomics, 77.9% for PET-based radiomics, and 72.1% for dosiomics based on the NN classifiers. The hybrid model exhibited the highest accuracy of 84.4% for CT-based radiomics based on the kNN classifier, 86.0% for PET-based radiomics, and 79.0% for dosiomics based on the NN classifiers. The proposed hybrid model exhibited promising predictive performance for the local response to definitive radiotherapy in ESCC patients. The prediction of the complete response for oesophageal cancer patients may contribute to improving overall survival. The hybrid model has the potential to improve prediction performance than the external validation model that was conventionally proposed. • Radiomics and dosiomics used to predict response in patients with oesophageal cancer receiving definitive radiotherapy. • Hybrid model with neural network classifier of PET-based radiomics improved prediction accuracy by 8.1%. • The hybrid model has the potential to improve prediction performance.

PCA 前处理对 KNN 识别猪脸的影响研究

To explore the application of traditional machine learning model in the intelligent management of pigs, in this paper, the influence of the PCA pre-treatment on pig face identification with KNN is studied. With testing method, individual identification test was carried out on 10 different pigs in two testing schemes, in which one adopted KNN alone and the other adopted PCA + KNN, for which the classifier parameter was taken as 3 and 5, respectively. In the optimized scheme, the operating efficiency got significantly increased, also the training time and testing time were reduced to 4.8% and 7% of the original value in the KNN alone scheme, though the accuracy got lowered to a certain extent. With all these factors taken into consideration, PCA pre-treatment is beneficial to individual pig identification with KNN. It can provide experimental support for mobile terminals and embedded application of KNN classifiers.

Detection of attention deficit hyperactivity disorder based on EEG signals using Least Square Support Vector Machine (LS-SVM)

ABSTRACT Attention Deficit Hyperactivity Disorder (ADHD) is a mental disorder affecting children in their early stages. Detection of ADHD is considered a challenging task for experts because there is no accurate test to determine whether a child has ADHD or not. Mainly, most of previous studies are used electroencephalography (EEG) signals to detect ADHD. In this study, a robust model for ADHD detection-integrated discrete wavelet transform (DWT), statistical features, and a least square support vector machine (LS-SVM) is proposed to detect ADHD from EEG signals. The EEG signals are decomposed into five bands and then, a set of statistical features are tested to find the optimal feature sets using K-means model. A public dataset is used to evaluate the proposed model. A total of 45 ADHD patients and 45 healthy are involved to evaluate the proposed model. Several metrics are used to evaluate the proposed model including 10-fold cross-validation precision, sensitivity, and specificity metrics. A channel selection is also investigated, and we found that the channels ‘P3, P7, Pz’ are more important and gave better prediction results than other channels. The proposed model obtained an average accuracy of 98.06% with LS-SVM, and 95.46% with the KNN classifier.

Extended right-angle difference ternary co-relation pattern: A new feature descriptor for face anti-spoofing

Presently, in many applications face recognition is commonly used as an authentication security system. However, an intruder can spoof the face recognition system by using genuine user photos or video to gain access. Therefore, there is a need for a superior technique for face anti-spoofing to handle all kinds of spoofing attacks in any environment. This paper presents a novel descriptor called extended right-angle difference ternary co-relation pattern (ERDTCP) for face anti-spoofing. The ERDTCP is calculated based on the division of a given image into 4 blocks at an angle of 0°, 90°, 180° and 270° from the center pixel. The effectiveness of the proposed method is evaluated on three publically available databases using KNN and SVM classifiers. Experiments were conducted using SOTA approaches and the proposed method using ERDTCP and the results were compared. The proposed face anti-spoofing approach performs better than SOTA approaches.

When Deep is not Enough: Towards Understanding Shallow and Continual Learning Models in Realistic Environmental Sound Classification for Robots

Although deep learning models are state-of-the-art models in audio classification, they fall short when applied in developmental robotic settings and human–robot interaction (HRI). The major drawback is that deep learning relies on supervised training with a large amount of data and annotations. In contrast, developmental learning strategies in human–robot interaction often deal with small-scale data acquired from HRI experiments and require the incremental addition of novel classes. Alternatively, shallow learning architectures that enable fast and yet robust learning are provided by simple distance metric-based learning and neural architectures implementing the reservoir computing paradigm. Similarly, continual learning algorithms receive more attention in the last years as they can integrate stable perceptual feature extraction using pre-trained deep learning models with open-set classification. As our research centers around reenacting the incremental learning of audio cues, we conducted a study on environmental sound classification using the iCaRL as well as the GDumb continual learning algorithms in comparison with a popular classifier in this domain, the kNN classifier, as well as employing an Echo State Network. We contrast our results with those obtained from a VGGish network that serves here as the performance upper bound that allows us to quantify the performance differences and to discuss current issues with continual learning in the audio domain. As only little is known about using shallow models or continual learning in the audio domain, we pass on additional techniques like data augmentation and create a simple experimental pipeline that is easy to reproduce. Although our selected algorithms are partially inferior in performance compared to the upper bound, our evaluation on three environmental sound datasets shows promising performance using continual learning for a subset of the DCASE2019 challenge dataset and the ESC10 dataset. As we do not address benchmarking in this paper, our study provides a good foundation for further research and computational improvements on shallow and continual learning models for robotic applications in the audio domain.