EEG Signal-Assisted Algebraic Topological Feature-Enhanced Deep Neural Networks for Gestalt Illusory Contour Perception

Deep neural networks (DNNs) have been incredibly successful at correctly classifying various input types, including images, speech, and data, according to consumer preferences. However, the efficacy of DNNs in Gestalt classification tasks, an important case in manifesting human perceptual capability, remains challenging, as DNNs are generally unable to perceive illusory closure from Gestalt images unless DNNs are carefully calibrated using a significant amount of priori information. However, altering the input, as minute changes typically imperceptible to humans, can confound carefully calibrated DNNs. In this study, EEG signal-empowered deep clustering based on topological data analysis (TDA) is proposed. The adoption of TDA manifests the separability of EEG signals responsive to Gestalt and non-Gestalt images. It also yields new families of features and descriptors for Gestalt illusory contours by extracting topological and geometric information. Furthermore, the combination of analyzed EEG signals and digital images further benefits the recognition of Gestalt illusory contours. Extensive experiments have shown convincing improvements over the state-of-the-art DNNs (e.g., DeepCluster and multi-view clustering [MVC]). In particular, DeepCluster with TDA can perceive illusory contours to some degree, given its 66.5% classification accuracy. Nevertheless, on top of extracted topological features from EEG signals, it produces higher classification accuracy (i.e., 71.9%), indicating the features extracted from the EEG signal contribute to perceive Gestalt illusory contours. On the other hand, the applications of topological features to the latest MVC also bring 4.5% improvement and demonstrate the effectiveness.