The human brain is the principal center for processing sensory information an‎d coordinating responses to the surrounding environment. However, neurological injuries an‎d disorders can severely disrupt these functions, leading to significant impairments in daily activities an‎d involuntary movements. Recent advances in brain–computer interface (BCI) technology aim to provide solutions for individuals suffering from these issues. Decoding visual information from neural signals—particularly in the face of high-dimensional data, limited sample sizes, an‎d the intrinsic noise of brain recordings—remains one of the main obstacles to developing sophisticated BCIs. To overcome these limitations, this dissertation introduces a novel nonparametric probabilistic generative framework called the Latent Discriminative Generative Decoder (LDGD), based on Gaussian processes. By jointly leveraging neural signal features an‎d visual stimulus labels, LDGD learns a shared low-dimensional latent space capable of both discriminative (classification) an‎d generative (data reconstruction) tasks. To enhance scalability an‎d enable mini-batch training, an inducing-point approximation is employed, an‎d a real-time variant, fastLDGD, is developed for online applications. The model was eva‎luated using intracranial electroencephalography (iEEG) data recorded from epilepsy patients performing two visual tasks: a simple binary stimulation task (flashing-checker) an‎d a complex multi-output task (10×10 checkerboard sequence). In the simple task, fastLDGD achieved an average accuracy of 72% outperforming stan‎dard machine-learning baselines (e.g. SVM an‎d XGBoost) an‎d deep-learning approaches (e.g., VAE an‎d EEGNet). In the complex checkerboard task, although uniform pixel-wise decoding did not exceed chance levels as expected, the model yielded important insights: it decoded the central pixel state with 69% accuracy, an‎d spatial accuracy analysis revealed a significant correlation between decoding performance an‎d visual attention, with accuracy decreasing as distance from the image center increased. This research demonstrates the efficacy of the proposed Bayesian framework for robustly decoding visual information from limited, noisy neural data. The findings an‎d models presented here represent an important step toward designing a new generation of interpretable BCIs for applications such as visual neuroprostheses an‎d functional brain mapping.

بهزاد نظري

علي يوسفي

محمدرضا احمدزاده , فرزانه شايق بروجني , علي مطيع نصرآبادي