The Impact of Mini-batch Design on EEG Classification in Anomaly Detection for Video Surveillance

In contemporary society, the exponential growth in video surveillance data has escalated the demand for automation in security and surveillance systems. Deep learning-based anomaly detection, commonly employed in these systems, presents a challenge in verifying whether the anomalies it identifies align with human perception of irregularities. As an initial attempt to bridge this gap, this study explores the utility of a system that detects anomalies based on electroencephalogram (EEG) recordings, which analyze the moments when humans recognize anomalies in video surveillance footage. A key finding pertains to the impact of mini-batch size and sequence on classification accuracy. Contrary to the conventional recommendation for classification problems (using random mini-batches of moderate size), our research discovered that employing small, non-randomized mini-batches enhances classification accuracy. Furthermore, centralizing the feature vector also improves accuracy. These three elements appear to contribute to domain adaptation effects. This underscores the importance of novel mini-batch designs in deep learning-based EEG classification problems that reflect the intrinsic structure and characteristics of EEG data, offering a new standardization method for data before employing complex models.