Abstract
Decoding visual object categories from electroencephalography (EEG) is a promising yet methodologically fragile problem. Reported performance is highly sensitive to dataset construction, preprocessing, evaluation protocols, and model design, making results difficult to interpret and compare across studies. This thesis adopts a validity-first perspective and investigates under which conditions multiclass visual object decoding from EEG can be considered reliable and meaningful.
First, a structured survey and controlled benchmarking framework (FOCUS) are used to analyze the methodological landscape of EEG-based visual decoding, revealing that performance is frequently confounded by block structure, stimulus repetition, and insufficiently specified splitting strategies. Second, a purpose-built dataset (EEG-CORD) is introduced to isolate protocol effects under confound-aware evaluation, showing that repetition, block design, session structure, and preprocessing choices can measurably influence decoding outcomes. Third, a multi-objective decoding framework (TRIAGE-EEG) is proposed to study the interaction between alignment-based EEG-image representation learning and explicit supervised classification. Experiments and ablations demonstrate that combining these objectives can improve performance in specific regimes, but that gains in these regimes also arise from architectural and optimization robustness.
Overall, the results show that decoding accuracy is not an intrinsic property of a model but an emergent consequence of methodological choices. The thesis provides practical recommendations for confound-aware evaluation and transparent reporting, emphasizing that sustained progress in EEG-based visual decoding depends on prioritizing validity and generalization over potentially fragile performance improvements.