ROC Analysis in Theory and Practice

Waste image classification based on transfer learning and convolutional neural network

The rapid economic and social development has led to a rapid increase in the output of domestic waste. How to realize waste classification through intelligent methods has become a key factor for human beings to achieve sustainable development. Traditional waste classification technology has low efficiency and low accuracy. To improve the efficiency and accuracy of waste classification processing, this paper proposes a DenseNet169 waste image classification model based on transfer learning. Because of the disadvantages of the existing public waste dataset, such as uneven distribution of data, single background, obvious features, and small sample size of the waste image, the waste image dataset NWNU-TRASH is constructed. The dataset has the advantages of balanced distribution, high diversity, and rich background, which is more in line with real needs. 70% of the dataset is used as the training set and 30% as the test set. Based on the deep learning network DenseNet169 pre-trained model, we can form a DenseNet169 model suitable for this experimental dataset. The experimental results show that the accuracy of classification is over 82% in the DenseNet169 model after the transfer learning, which is better than other image classification algorithms.

Testing encoding specificity and the diagnostic feature-detection theory of eyewitness identification, with implications for showups, lineups, and partially disguised perpetrators

The diagnostic feature-detection theory (DFT) of eyewitness identification is based on facial information that is diagnostic versus non-diagnostic of suspect guilt. It primarily has been tested by discounting non-diagnostic information at retrieval, typically by surrounding a single suspect showup with good fillers to create a lineup. We tested additional DFT predictions by manipulating the presence of facial information (i.e., the exterior region of the face) at both encoding and retrieval with a large between-subjects factorial design (N = 19,414). In support of DFT and in replication of the literature, lineups yielded higher discriminability than showups. In support of encoding specificity, conditions that matched information between encoding and retrieval were generally superior to mismatch conditions. More importantly, we supported several DFT and encoding specificity predictions not previously tested, including that (a) adding non-diagnostic information will reduce discriminability for showups more so than lineups, and (b) removing diagnostic information will lower discriminability for both showups and lineups. These results have implications for police deciding whether to conduct a showup or a lineup, and when dealing with partially disguised perpetrators (e.g., wearing a hoodie).

A perceptual scaling approach to eyewitness identification

Eyewitness misidentification accounts for 70% of verified erroneous convictions. To address this alarming phenomenon, research has focused on factors that influence likelihood of correct identification, such as the manner in which a lineup is conducted. Traditional lineups rely on overt eyewitness responses that confound two covert factors: strength of recognition memory and the criterion for deciding what memory strength is sufficient for identification. Here we describe a lineup that permits estimation of memory strength independent of decision criterion. Our procedure employs powerful techniques developed in studies of perception and memory: perceptual scaling and signal detection analysis. Using these tools, we scale memory strengths elicited by lineup faces, and quantify performance of a binary classifier tasked with distinguishing perpetrator from innocent suspect. This approach reveals structure of memory inaccessible using traditional lineups and renders accurate identifications uninfluenced by decision bias. The approach furthermore yields a quantitative index of individual eyewitness performance.

VR motion sickness recognition by using EEG rhythm energy ratio based on wavelet packet transform

BACKGROUND AND OBJECTIVES

Virtual reality motion sickness (VRMS) is one of the main factors hindering the development of VR technology. At present, the VRMS recognition methods using electroencephalogram (EEG) signals have poor applicability to multiple subjects.

METHODS

Aiming at this dilemma, the wavelet packet transform (WPT), was used to propose a feature extraction method for EEG rhythm energy ratios of delta (δ), theta (θ), alpha (α), and beta (β) in this research. Moreover, VRMS was recognized by combining k-Nearest Neighbor classifier (k-NN), support vector machine (SVM) with polynomial kernel (polynomial-SVM) and radial basis function kernel (RBF-SVM), respectively. The method is that the raw EEG signals were de-noised by an elliptical band-pass filter and segmented by a fixed window, 7-level db4 WPT was performed on each EEG segment, and the wavelet packet energy ratios of delta, theta, alpha and beta rhythms from FP1, FP2, C3, C4, P3, P4, O1 and O2 channels were calculated and combined to form feature vectors for recognizing VRMS.

RESULTS

Under the condition of 4-s window size, the average VRMS recognition accuracy of polynomial-SVM for the single subject was 92.85%, and the VRMS recognition accuracy of 18 subjects was about 79.25%.

CONCLUSIONS

Compared with other VRMS recognition methods, this method does not only have a higher recognition accuracy to a single subject, but also have better applicability to multiple subjects. Meanwhile, when using the EEG four rhythm energy ratios of FP1, FP2, C3, C4, P3, P4, O1 and O2 channels as feature vectors, the polynomial-SVM achieved better VRMS recognition performance than the k-NN and RBF-SVM.

Mitigating cue competition effects in human category learning

When people learn perceptual categories, if one feature makes it easy to determine the category membership, learning about other features can be reduced. In three experiments, we asked whether this cue competition effect could be fully eradicated with simple instructions. For this purpose, in a pilot experiment, we adapted a classical overshadowing paradigm into a human category learning task. Unlike previous reports, we demonstrate a robust cue competition effect with human learners. In Experiments 1 and 2, we created a new warning condition that aimed at eradicating the cue competition effect through top-down instructions. With a medium-size overshadowing effect, Experiment 1 shows a weak mitigation of the overshadowing effect. We replaced the stimuli in Experiment 2 to obtain a larger overshadowing effect and showed a larger warning effect. Nevertheless, the overshadowing effect could not be fully eradicated. These experiments suggest that cue competition effects can be a stubborn roadblock in human category learning. Theoretical and practical implications are discussed.

Models of lineup memory

Face recognition memory is often tested by the police using a photo lineup, which consists of one suspect, who is either innocent or guilty, and five or more physically similar fillers, all of whom are known to be innocent. For many years, lineups were investigated in lab studies without guidance from standard models of recognition memory. More recently, signal detection theory has been used to conceptualize lineup memory and to motivate receiver operating characteristic (ROC) analysis of lineup performance. Here, we describe three competing signal-detection models of lineup memory, derive their likelihood functions, and fit them to empirical ROC data. We also introduce the notion that memory signals generated by the faces in a lineup are likely to be correlated because, by design, those faces share features. The models we investigate differ in their predictions about the effect that correlated memory signals should have on the ability to discriminate innocent from guilty suspects. A popular compound signal detection model known as the Integration model predicts that correlated memory signals should impair discriminability. Empirically, this model performed so poorly that, going forward, it should probably be abandoned. The best-fitting model incorporates a principle known as "ensemble coding," which predicts that correlated memory signals should enhance discriminability. The ensemble model aligns with a previously proposed theory of eyewitness identification according to which the simultaneous presentation of faces in a lineup enhances discriminability compared to when faces are presented in isolation because it permits eyewitnesses to detect and discount non-diagnostic facial features.

Theoretical vs. empirical discriminability: the application of ROC methods to eyewitness identification

Receiver operating characteristic (ROC) analysis was introduced to the field of eyewitness identification 5 years ago. Since that time, it has been both influential and controversial, and the debate has raised an issue about measuring discriminability that is rarely considered. The issue concerns the distinction between empirical discriminability (measured by area under the ROC curve) vs. underlying/theoretical discriminability (measured by d’ or variants of it). Under most circumstances, the two measures will agree about a difference between two conditions in terms of discriminability. However, it is possible for them to disagree, and that fact can lead to confusion about which condition actually yields higher discriminability. For example, if the two conditions have implications for real-world practice (e.g., a comparison of competing lineup formats), should a policymaker rely on the area-under-the-curve measure or the theory-based measure? Here, we illustrate the fact that a given empirical ROC yields as many underlying discriminability measures as there are theories that one is willing to take seriously. No matter which theory is correct, for practical purposes, the singular area-under-the-curve measure best identifies the diagnostically superior procedure. For that reason, area under the ROC curve informs policy in a way that underlying theoretical discriminability never can. At the same time, theoretical measures of discriminability are equally important, but for a different reason. Without an adequate theoretical understanding of the relevant task, the field will be in no position to enhance empirical discriminability.