Degrade Classification Accuracy Research Articles

A novel feature selection method based on quantum support vector machine

Feature selection is critical in machine learning to reduce dimensionality and improve model accuracy and efficiency. The exponential growth in feature space dimensionality for modern datasets directly results in ambiguous samples and redundant features, which can severely degrade classification accuracy. Quantum machine learning offers potential advantages for addressing this challenge. In this paper, we propose a novel method, quantum support vector machine feature selection (QSVMF), integrating quantum support vector machines with multi-objective genetic algorithm. QSVMF optimizes multiple simultaneous objectives: maximizing classification accuracy, minimizing selected features and quantum circuit costs, and reducing feature covariance. We apply QSVMF for feature selection on a breast cancer dataset, comparing the performance of QSVMF against classical approaches with the selected features. Experimental results show that QSVMF achieves superior performance. Furthermore, the Pareto front solutions of QSVMF enable analysis of accuracy versus feature set size trade-offs, identifying extremely sparse yet accurate feature subsets. We contextualize the biological relevance of the selected features in terms of known breast cancer biomarkers. This work highlights the potential of quantum-based feature selection to enhance machine learning efficiency and performance on complex real-world data.

MT Method for Anomaly Detection and Classification using EM-λ Algorithm

Purpose: In this paper, we propose a method to classify and detect normal, known anomalies, and unknown anomalies by combining the expectation–maximisation (EM-λ) algorithm and the Mahalanobis–Taguchi (MT) method. Methodology/Approach: The proposed method learns normal data that are expected to be homogeneous and known abnormal data and performs classification and detection by parameter estimation using the EM-λ algorithm. Conventional methods perform analysis based on parameter estimation using the EM algorithm. However, the EM algorithm can degrade classification accuracy if it does not assume that the data fits the model's generative process. Findings: We verify the performance of the proposed method using artificially generated data and real-world bean data for classification as data that do not satisfy this assumption. The validation results show up to 6% improvement over the conventional method in classification accuracy and unknown anomaly discrimination accuracy. Research Limitation/implication: We try various patterns for the parameter of the proposed method in the verification. However, this way is computationally expensive. Originality/Value of paper: Conventional methods perform analysis based on parameter estimation using the EM algorithm. Our proposal method seeks to improve accuracy by using the EM-λ algorithm for parameter estimation, which is expected to improve classification accuracy when the data do not conform to the generative assumptions of the EM algorithm's model.

A Deep Neural Network Attack Simulation against Data Storage of Autonomous Vehicles

<div>In the pursuit of advancing autonomous vehicles (AVs), data-driven algorithms have become pivotal in replacing human perception and decision-making. While deep neural networks (DNNs) hold promise for perception tasks, the potential for catastrophic consequences due to algorithmic flaws is concerning. A well-known incident in 2016, involving a Tesla autopilot misidentifying a white truck as a cloud, underscores the risks and security vulnerabilities. In this article, we present a novel threat model and risk assessment (TARA) analysis on AV data storage, delving into potential threats and damage scenarios. Specifically, we focus on DNN parameter manipulation attacks, evaluating their impact on three distinct algorithms for traffic sign classification and lane assist. Our comprehensive tests and simulations reveal that even a single bit-flip of a DNN parameter can severely degrade classification accuracy to less than 10%, posing significant risks to the overall performance and safety of AVs. Additionally, we identify critical parameters based on bit position, layer position, and bit-flipping direction, offering essential insights for developing robust security measures in autonomous vehicle systems.</div>

Ergonomic risk level prediction framework for multiclass imbalanced data

Work-related Musculoskeletal Disorders (WMSDs) are a common concern in the manufacturing industry and are mostly induced by postural load requirements of job tasks that affect the neck, trunk, and upper extremities. Ergonomics assessment, such as Rapid Upper Limb Assessment (RULA), and risk prediction can help industries prevent and evaluate the risk of WMSDs. The real-world practice consists of multiple human activities that cause difficulties to assess with a common learning approach. In particular, the issue of imbalanced class distributions can degrade classification accuracy significantly. This study aims to propose a three-stage framework for predicting the risk of WMSDs that handles imbalanced data. The first stage is data preparation which consists of collecting and processing data. The second stage is the evaluation of our RULA score calculation. In the last stage, we predict the risk level in two steps. The first step is to oversample the imbalanced class with an extended approach by considering relevant activities, called KD-SMOTE. The second step is to predict the risk level using our proposed network, called HyNet-CB, which is a classifier based on hybrid neural network architecture that combines Convolutional Neural Network and Bidirectional Long Short-Term Memory (HyNet-CB). We evaluate the framework using a collected dataset of an assembly process. Our framework achieved a 92.43% F1-score, and it can classify all classes with more than 80% F1-score.

Leveraging Tactile Sensors for Low Latency Embedded Smart Hands for Prosthetic and Robotic Applications

Tactile sensing is a crucial perception mode for robots and human amputees in need of controlling a prosthetic device. Today, robotic and prosthetic systems are still missing the important feature of accurate tactile sensing. This lack is mainly due to the fact that the existing tactile technologies have limited spatial and temporal resolution and are either expensive or not scalable. In this article, we present the design and implementation of a hardware–software embedded system called SmartHand. It is specifically designed to enable the acquisition and real-time processing of high-resolution tactile information from a hand-shaped multisensor array for prosthetic and robotic applications. During data collection, our system can deliver a high throughput of 100 frames per second, which is <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$13.7\times $ </tex-math></inline-formula> higher than previous related work. This has allowed the collection of a new tactile dataset consisting of 340 000 frames while interacting with 16 objects from everyday life during five different sessions. Together with the empty hand, the dataset presents a total of 17 classes. We propose a compact yet accurate convolutional neural network that requires one order of magnitude less memory and <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$15.6\times $ </tex-math></inline-formula> fewer computations compared with related work without degrading classification accuracy. The top-1 and top-3 cross-validation accuracies on the collected dataset are, respectively, 98.86% and 99.83%. We further analyze the intersession variability and obtain the best top-3 leave-one-out-validation accuracy of 77.84%. We deploy the trained model on a high-performance ARM Cortex-M7 microcontroller achieving an inference time of only 100 ms minimizing the response latency. The overall measured power consumption is 505 mW. Finally, we fabricate a new control sensor and perform additional experiments to provide analyses on sensor degradation and slip detection. This work is a step forward in giving robotic and prosthetic devices a sense of touch by demonstrating the practicality of a smart embedded system that uses a scalable tactile sensor with embedded tiny machine learning.

A Full Tensor Network for Hyperspectral Image Classification

Hyperspectral image (HSI) is widely used in real-world classification tasks since it contains rich spatial and spectral features consisting of hundreds of continuous bands. However, the rich spectral features always result in “dimension disaster” for HSI classification. Some traditional methods adopt the strategy of dimensionality reduction in advance to alleviate the influences of high spectral feature dimension, which usually results in degraded classification accuracy due to the hard coupling of the dimensionality reduction model and the classifier. In this letter, we propose a full tensor network for HSI classification based on Tucker decomposition-based hidden layer and CP decomposition-based classification layer. The proposed network can efficiently extract the high-level structural features directly from the original data and classify HSI in a single network without dimension reduction procedure in advance. Moreover, the proposed model has a simple structure and holds less parameters than the state-of-art classification methods. The experimental results on Pavia University and Salinas datasets show that for training ratio of 5%, the overall accuracy of our model are averaged 3.41% higher than CNN models while the network parameters are averaged 80.44% less, which validate the merits of the proposed method.

MASKER: Masked Keyword Regularization for Reliable Text Classification

Pre-trained language models have achieved state-of-the-art accuracies on various text classification tasks, e.g., sentiment analysis, natural language inference, and semantic textual similarity. However, the reliability of the fine-tuned text classifiers is an often underlooked performance criterion. For instance, one may desire a model that can detect out-of-distribution (OOD) samples (drawn far from training distribution) or be robust against domain shifts. We claim that one central obstacle to the reliability is the over-reliance of the model on a limited number of keywords, instead of looking at the whole context. In particular, we find that (a) OOD samples often contain in-distribution keywords, while (b) cross-domain samples may not always contain keywords; over-relying on the keywords can be problematic for both cases. In light of this observation, we propose a simple yet effective fine-tuning method, coined masked keyword regularization (MASKER), that facilitates context-based prediction. MASKER regularizes the model to reconstruct the keywords from the rest of the words and make low-confidence predictions without enough context. When applied to various pre-trained language models (e.g., BERT, RoBERTa, and ALBERT), we demonstrate that MASKER improves OOD detection and cross-domain generalization without degrading classification accuracy. Code is available at https://github.com/alinlab/MASKER.

AresB-Net: accurate residual binarized neural networks using shortcut concatenation and shuffled grouped convolution.

This article proposes a novel network model to achieve better accurate residual binarized convolutional neural networks (CNNs), denoted as AresB-Net. Even though residual CNNs enhance the classification accuracy of binarized neural networks with increasing feature resolution, the degraded classification accuracy is still the primary concern compared with real-valued residual CNNs. AresB-Net consists of novel basic blocks to amortize the severe error from the binarization, suggesting a well-balanced pyramid structure without downsampling convolution. In each basic block, the shortcut is added to the convolution output and then concatenated, and then the expanded channels are shuffled for the next grouped convolution. In the downsampling when stride >1, our model adopts only the max-pooling layer for generating low-cost shortcut. This structure facilitates the feature reuse from the previous layers, thus alleviating the error from the binarized convolution and increasing the classification accuracy with reduced computational costs and small weight storage requirements. Despite low hardware costs from the binarized computations, the proposed model achieves remarkable classification accuracies on the CIFAR and ImageNet datasets.

Detection of Backdoors in Trained Classifiers Without Access to the Training Set.

With wide deployment of deep neural network (DNN) classifiers, there is great potential for harm from adversarial learning attacks. Recently, a special type of data poisoning (DP) attack, known as a backdoor (or Trojan), was proposed. These attacks do not seek to degrade classification accuracy, but rather to have the classifier learn to classify to a target class t∗ whenever the backdoor pattern is present in a test example originally from a source class s∗ . Launching backdoor attacks does not require knowledge of the classifier or its training process-only the ability to poison the training set with exemplars containing a backdoor pattern (labeled with the target class). Defenses against backdoors can be deployed before/during training, post-training, or at test time. Here, we address post-training detection in DNN image classifiers, seldom considered in existing works, wherein the defender does not have access to the poisoned training set, but only to the trained classifier itself, as well as to clean (unpoisoned) examples from the classification domain. This scenario is of great interest because e.g., a classifier may be the basis of a phone app that will be shared with many users. Detection may thus reveal a widespread attack. We propose a purely unsupervised anomaly detection (AD) defense against imperceptible backdoor attacks that: 1) detects whether the trained DNN has been backdoor-attacked; 2) infers the source and target classes in a detected attack; 3) estimates the backdoor pattern itself. Our AD approach involves learning (via suitable cost function minimization) the minimum size/norm perturbation (putative backdoor) required to induce the classifier to misclassify (most) examples from class s to class t , for all (s,t) pairs. Our hypothesis is that nonattacked pairs require large perturbations, while the attacked pair (s∗, t∗) requires much smaller ones. This is convincingly borne out experimentally. We identify a variety of plausible cost functions and devise a novel, robust hypothesis testing approach to perform detection inference. We test our approach, in comparison with the state-of-the-art methods, for several backdoor patterns, attack settings and mechanisms, and data sets and demonstrate its favorability. Our defense essentially requires setting a single hyperparameter (the detection threshold), which can e.g., be chosen to fix the system's false positive rate.

A Hybrid Approach of Dimensionality Reduction and Classification for Hyperspectral Imager

Hyperspectral imagery widely used in remote sensing applications provides detailed data for the sensed materials than multispectral imagery. In this image, the identification of every ground surface picture element is represented by its corresponding spectral signatures which are unique but difficult to process as a result of the large volume of information. Traditional classification strategies might not be used along with the dimension reduction technique. This is often as a result of large dimension, that refers to the very fact that the sample size required to estimate a perform of many variables to a given degree of accuracy grows exponentially with the quantity of variables. Principal Component Analysis (PCA) has been the technique of alternative for dimension reduction. However, PCA is computationally extensive and not eliminate anomalies which will be seen at an arbitrary band. The high dimensional nature of the information collected by such sensors not solely will increase procedure complexness however can also degrade classification accuracy. To deal with this issue, dimension reduction (DR) has become a crucial aid to rising classifier potency on these imagery. In this paper, we summarized the hybrid approach of dimensionality reduction of HSI for the purpose of effective classification. The key findings revealed that the classification accuracy has been increased significantly on the application of dimensionality reduction techniques.

A Refinement: Better Classification of Images using LDA in Contrast with SURF and SVM for CBIR System

Content Based Image Retrieval or CBIR is the image retrieval process which is based on visual features including color, texture and shape. Image databases that are large in size and traditional indexing of images have proven to be less sufficient, more laborious and extremely time consuming which led to its development. Image retrieval and face recognition systems have grown significantly in the area of security systems. Retrieval of images has become an challenging issue in real world applications due to high dimensions of image collection. Therefore it has become a major issue and downfall for Content-Based image Retrieval (CBIR). Furthermore, it results in the inefficiency and degraded classification accuracy. A classifier in combination with the better dimensionality reduction and higher class discrimination can provide higher classification accuracy. This combination must result in a better classification rate. In the field of medical image annotation, research shows that SURF is a very strong tool to be used. Herein, an effort has been formulated to present an efficient algorithm based on SURF which is fast and robust interest point detector, SVM and LDA for further classification. General Terms Feature Extraction, Description, Detection, Classification, Image Retrieval, Dimensionality Reduction.

Integrating geospatial and multi‐depth laboratory spectral data for mapping soil classes in a geologically complex area in southeastern Brazil

SummarySoil mapping across large areas can be enhanced by integrating different methods and data sources. This study merges laboratory, field and remote sensing data to create digital maps of soil suborders based on the Brazilian Soil Classification System, with and without additional textural classification, in an area of 13 000 ha in the state of São Paulo, southeastern Brazil. Data from 289 visited soil profiles were used in multinomial logistic regression to predict soil suborders from geospatial data (geology, topography, emissivity and vegetation index) and visible–near infrared (400–2500 nm) reflectance of soil samples collected at three depths (0–20, 40–60 and 80–100 cm). The derived maps were validated with 47 external observations, and compared with two conventional soil maps at scales of 1:100 000 and 1:20 000. Soil suborders with and without textural classification were predicted correctly for 44 and 52% of the soil profiles, respectively. The derived suborder maps agreed with the 1:100 000 and 1:20 000 conventional maps in 20 and 23% (with textural classification) and 41 and 46% (without textural classification) of the area, respectively. Soils that were well defined along relief gradients (Latosols and Argisols) were predicted with up to 91% agreement, whereas soils in complex areas (Cambisols and Neosols) were poorly predicted. Adding textural classification to suborders considerably degraded classification accuracy; thus modelling at the suborder level alone is recommended. Stream density and laboratory soil reflectance improved all classification models, showing their potential to aid digital soil mapping in complex tropical environments.

유도전동기의 고장 진단을 위한 효과적인 특징 추출 방법

본 논문은 고장 분류 시스템을 위해 진동 신호로부터 특징 벡터를 자동적으로 추출하는 효과적인 기법을 제안한다. 기존의 멜-주파수 캡스트럼 계수는 진동신호의 노이즈에 민감하여 분류 정확도를 감소시키는 단점이 있다. 이러한 문제를 해결하기 위해 본 논문은 4단계 필터 뱅크로 구성된 스펙트럴 엔벨로프 캡스트럼 계수 분석을 제안하며, 4단계는 (1) 모든 진동 신호의 스펙트럴 엔벨로프를 기술하기 위한 선형 예측 코딩 알고리즘 사용 단계, (2) 일반적인 스펙트럴 모양을 얻기 위해 모든 엔벨로프의 평균화 단계, (3) 평균 엔벨로프와 그 주파수의 최대값을 찾기 위한 기울기 하강 방법 사용 단계, (4) 엔벨로프의 주파수 사이의 거리로부터 계산된 중앙값을 얻는데 사용되는 비 중첩 필터 뱅크 단계로 구성된다. 이4-단계필터뱅크는 특징벡터를 추출하기위해 캡스트럼 계수 계산에 사용된다. 마지막으로 유도전동기의 결함 형태를 구분하기 위해 이러한 특수 파라미터를 사용하는 다중 계층 서포트 벡터 머신을 사용한다. 모의실험 결과, 제안하는 방법은 약 99.65%의 분류 성능을 보이며, 동시에 기존 방법들보다 우수한 성능을 보인다. This paper proposes an effective technique that is used to automatically extract feature vectors from vibration signals for fault classification systems. Conventional mel-frequency cepstral coefficients (MFCCs) are sensitive to noise of vibration signals, degrading classification accuracy. To solve this problem, this paper proposes spectral envelope cepstral coefficients (SECC) analysis, where a 4-step filter bank based on spectral envelopes of vibration signals is used: (1) a linear predictive coding (LPC) algorithm is used to specify spectral envelopes of all faulty vibration signals, (2) all envelopes are averaged to get general spectral shape, (3) a gradient descent method is used to find extremes of the average envelope and its frequencies, (4) a non-overlapped filter is used to have centers calculated from distances between valley frequencies of the envelope. This 4-step filter bank is then used in cepstral coefficients computation to extract feature vectors. Finally, a multi-layer support vector machine (MLSVM) with various sigma values uses these special parameters to identify faulty types of induction motors. Experimental results indicate that the proposed extraction method outperforms other feature extraction algorithms, yielding more than about 99.65% of classification accuracy.

Comparative Study of Intrinsic Dimensionality Estimation and Dimension Reduction Techniques on Hyperspectral Images Using K-NN Classifier

Nowadays, hyperspectral remote sensors are readily available for monitoring the Earth's surface with high spectral resolution. The high-dimensional nature of the data collected by such sensors not only increases computational complexity but also can degrade classification accuracy. To address this issue, dimensionality reduction (DR) has become an important aid to improving classifier efficiency on these images. The common approach to decreasing dimensionality is feature extraction by considering the intrinsic dimensionality (ID) of the data. A wide range of techniques for ID estimation (IDE) and DR for hyperspectral images have been presented in the literature. However, the most effective and optimum methods for IDE and DR have not been determined for hyperspectral sensors, and this causes ambiguity in selecting the appropriate techniques for processing hyperspectral images. In this letter, we discuss and compare ten IDE and six DR methods in order to investigate and compare their performance for the purpose of supervised hyperspectral image classification by using K-nearest neighbor (K-NN). Due to the nature of K-NN classifier that uses different distance metrics, a variety of distance metrics were used and compared in this procedure. This letter presents a review and comparative study of techniques used for IDE and DR and identifies the best methods for IDE and DR in the context of hyperspectral image analysis. The results clearly show the superiority of the hyperspectral signal subspace identification by minimum, second moment linear, and noise-whitened Harsanyi-Farrand-Chang estimators, also the principal component analysis and independent component analysis as DR techniques, and the norm L1 and Euclidean distance metrics to process hyperspectral imagery by using the K-NN classifier.

Training Set Compression by Incremental Clustering

Abstract Compression of training sets is a technique for reducing training set size without degrading classifi-cation accuracy. By reducing the size of a training set, training will be more efficient in addition tosaving storage space. In this paper, an incremental clustering algorithm, the Leader algorithm, is usedto reduce the size of a training set by effectively subsampling the training set. Experiments on sev-eral standard data sets using SVM and KNN as classifiers indic ate that the proposed method is moreefficient than CONDENSE in reducing the size of training set w ithout degrading the classificationaccuracy. While the compression ratio for the CONDENSE method is fixed, the proposed methodoffers variable compression ratio through the cluster threshold value.Keywords: Clustering, Support vector machine, KNN, Pattern recognition, CONDENSE. 1. Introduction The training and/or testing complexity of a classifier usual ly depends on the size of the trainingset, e.g. the nearest neighbor (NN) classifier [1]. Nearest n eighbor and its generalized form, theK-nearest neighbor (KNN) classifier, are among the most popular non-parametric classifiers. Themembership of an unknown sample is classified based on the maj ority vote of the K nearest neigh-bors. There is no explicit learning from the training set. The entire training set itself defines thedecision boundaries. It is conceptually simple and shows good performance in many applications,e.g. it was used in face recognition for visitor identificati on [2] and it was shown that it outper-formed more sophisticated algorithms that use Principal Components Analysis (PCA) and neuralnetworks. Unfortunately, when the size of the training set is high, it requires a lot of memory tostore the entire training set and it also takes longer to search for the nearest neighbors of a given testpattern to make a single membership classification. Obvious ly, reducing the size of a training setcan improve the space and time efficiency of KNN. There has bee n considerable interest in reducingthe training set size by editing, especially in the context of NN. Different proximity graphs (such asDelaunay triangulation) may be used for editing NN rules [3,4]. Complexities of such approachesare prohibitively high. For example, the Voronoi diagram has worst case complexity ofΘ(n

A study of regularized Gaussian classifier in high-dimension small sample set case based on MDL principle with application to spectrum recognition

In classifying high-dimensional patterns such as stellar spectra by a Gaussian classifier, the covariance matrix estimated with a small-number sample set becomes unstable, leading to degraded classification accuracy. In this paper, we investigate the covariance matrix estimation problem for small-number samples with high dimension setting based on minimum description length (MDL) principle. A new covariance matrix estimator is developed, and a formula for fast estimation of regularization parameters is derived. Experiments on spectrum pattern recognition are conducted to investigate the classification accuracy with the developed covariance matrix estimator. Higher classification accuracy results are obtained and demonstrated in our new approach.

Sensor Selection for Machine Olfaction Based on Transient Feature Extraction

Machine olfaction devices, which are often called electronic noses (e-noses), are gaining favor for odor assessment applications in several industrial sectors, such as beverage, perfumery, and food. From a design point of view, the number of sensors in these devices for a particular odor application should be minimized without degrading classification accuracy. This paper deals with selecting sensors for e-noses to make small portable devices with fast response times and reduced cost possible. Prior research efforts have been reported in the open literature and have shown that many advantages can be gained by properly selecting the input features before forwarding to a pattern classification algorithm. This selection process can reduce the dimensionality of the feature space, remove redundant and irrelevant features, speed up classification, and improve classification performance. In this paper, the transient features of an array of sensors obtained by applying a multiresolutional approximation technique from the discrete wavelet transform (DWT) are investigated to search for an optimal sensor array to be implemented in the e-nose system. A genetic algorithm is adapted to tailor a gas sensor array for two different odor data sets (coffee and soda). From the experimental results, the input features obtained by applying the DWT to the transient sensor responses not only provide a significant reduction in the number of sensors when compared to traditional features but also improve the classification rate to near 100%.

The effect of principal component analysis on machine learning accuracy with high-dimensional spectral data

This paper presents the results of an investigation into the use of machine learning methods for the identification of narcotics from Raman spectra. The classification of spectral data and other high-dimensional data, such as images, gene-expression data and spectral data, poses an interesting challenge to machine learning, as the presence of high numbers of redundant or highly correlated attributes can seriously degrade classification accuracy. This paper investigates the use of principal component analysis (PCA) to reduce high-dimensional spectral data and to improve the predictive performance of some well-known machine learning methods. Experiments are carried out on a high-dimensional spectral dataset. These experiments employ the NIPALS (Non-Linear Iterative Partial Least Squares) PCA method, a method that has been used in the field of chemometrics for spectral classification, and is a more efficient alternative than the widely used eigenvector decomposition approach. The experiments show that the use of this PCA method can improve the performance of machine learning in the classification of high-dimensional data.

On pattern recognition dependency of desorption heat, activation energy, and temperature of polymer-based VOC sensors for the electronic NOSE

Design of polymer-based, chemo-responsive chemical sensors for a given set of target vapors is a challenging task, as the detailed mechanism of the sensing phenomena is still not well understood. In this work, we propose a novel approach to select gas-sensing materials for a given set of volatile organic compounds (VOC) when the governing phenomenon that drives gas/vapor interaction is adsorption. We do so by varying the activation energy, E A; enthalpy of adsorption, Q; and temperature, T, that directly relate to gas–solid interaction followed by neural network pattern recognition to evaluate to what extent variations in adsorption energetics and operational temperature results in satisfactory or unsatisfactory adsorbate classification. The adsorption of four gases; toluene, benzene, heptane and butanol at six discrete polymer sensors was simulated by varying E A and Q for different gas–solid interaction at normal temperature and pressure (NTP) and by varying T for a fixed set of E A and Q conditions. Data sets thus generated were used to build a feed-forward neural network (NN) model to classify the vapors. Performance of the NN model was evaluated by its ability to correctly identify the various gases. Pattern recognition dependency upon variation of activation energy and desorption heat was evaluated using data generated with perturbed E A. Upper and lower bound variation of the activation energy were found in order to retain within acceptable classification accuracy. It has been found that to be able to remain a classification accuracy of at least 90%, the tolerated perturbation of the activation energy should not exceed, in absolute value, 3% of that used during the training step of the neural networks. Similarly, temperature shifts that are set outside the range −4% to +3% of the training temperature were found to degrade classification accuracy to below acceptable limits of 90%. These results should be taken into consideration when designing and fabricating chemoresistive polymer sensors for use in an e-natural olfactory sensor emulator (e-NOSE™) ANN system.

Hard and soft classifications by a neural network with a non-exhaustively defined set of classes

Many assumptions are typically made in the course of a supervised digital image classification. The focus of this paper is the commonly made assumption of an exhaustively defined set of classes. This assumption is often unsatisfied, with the imagery containing regions of classes that were not included in training the classification. The failure to satisfy the assumed condition was investigated with reference to hard and soft land cover classifications by a feedforward neural network. The accuracy of these classifications was decreased if a non-exhaustively defined set of classes was used. The exclusion of a class from the training stage resulted in a decrease in the accuracy of a hard classification of agricultural crops of up to 21.2%. Moreover, there were marked differences between the 'real' and 'apparent' accuracies of classifications of up to 15.4%. With soft classifications of urban land cover, the presence of an untrained class also degraded classification accuracy. In the soft classification output, the correlation between the actual and estimated proportional cover of a class declined from r =0.97 to r =0.76 when another class was excluded from the training stage of the classification. Possible means to reduce the negative impacts of untrained classes are considered briefly. Post-classification thresholding of the neural network's output unit activation levels may form the basis of a method to identify and remove cases of an untrained class from a hard classification. Alternatively, supporting information on the typicality of class membership may be used to identify cases representing an area containing an untrained class in both hard and soft classifications and this is illustrated with reference to a soft classification.