Adaptive resonance clustering for out-of-distribution detection and its application to human-computer interaction

Abstract

Out-of-distribution detection is one of the most critical problems in human-computer interaction. Various application methods have been tried to solve data classification problems using deep-learning-based classification networks. However, most supervised classification networks are trained using softmax cross-entropy to achieve high classification performance. Then the networks often misclassify the out-of-distribution inputs as one of the known classification labels with high confidence, resulting in an overconfidence problem. When the overconfidence problem occurs, the model silently fails with no alarm to out-of-distribution inputs. It is a factor that seriously affects the reliability of the model from the point of view of real-world application. Out-of-distribution detection is one way to mitigate the side effect of declining model reliability in the overconfidence problem. Related studies that estimate a confidence score using classification networks have been proposed by adopting maximum softmax probability, cosine similarity, Mahalanobis distance, generative models, self-supervised learning, exposing outliers, or minimizing an empirical risk. Many out-of-distribution detection techniques have shown promising performance not only in the visual domain but also in the non-visual domain. Instead of directly classifying the out-of-distribution input, they bypass the decision threshold determination by estimating or calculating the degree of anomaly confidence score. Therefore, the confidence score-based out-of-distribution detection techniques require the user to determine a decision threshold to classify the out-of-distribution input directly. The primary method for finding the decision threshold is the outlier exposure method to adjust the decision threshold, which determines the threshold with the value that shows the best performance using both the in-distribution and out-of-distribution datasets. This property violates the closed-world assumption that deep-learning classification models cannot know anomalies during the training and deployment phases. The dissertation proposes a new framework to explore the decision threshold margin using only in-distribution data and to binary classify out-of-distribution data directly. The proposed framework adopts the resonance clustering method to detect out-of-distribution. The resonance clusters are trained by using cosine similarity between the concept vector of the cluster and preprocessed input features from a pre-trained feature extraction network. The proposed framework directly classifies in-distribution and out-of-distribution samples without outlier exposure by using clusters that are only trained with in-distribution data samples. We quantitatively compare the performance of the proposed framework and related studies using open visual datasets. We apply the proposed framework to two real-time applications to show the validity of the proposed framework: Frequency modulated continuous wave radar input and RGB camera input-based real-time hand gesture recognition systems capable of out-of-distribution detection. This study introduces a series of out-of-distribution detection processes from data collection for clustering to inference. We reported the out-of-distribution detection performance of the proposed system. The proposed method shows potential as a real-world application through real-time experimental demonstration. We performed outlier exposure to the proposed framework by assuming cases where out-of-distribution data can be obtained. It is confirmed that the performance of the proposed system can be significantly improved by applying the outlier exposure technique. Finally, through real-time experimental demonstrations, it is confirmed that it is feasible to use the proposed framework as a real-world application.