EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models
Researchers have introduced EEG-FM-Bench, a new benchmark designed to standardize the evaluation of electroencephalography foundation models (EEG-FMs). This benchmark integrates 14 datasets and 10 paradigms, incorporating various fine-tuning strategies and analysis tools. Initial experiments suggest that while multi-task learning can act as a regularizer, pre-training efficiency is hindered by gradient conflicts. The findings also indicate that model or data scale alone does not fully determine transfer performance, with objective alignment and adaptation compatibility playing crucial roles. AI