Are EEG-to-Text Models Working?

TL;DR

This paper critically evaluates EEG-to-Text models, revealing that current evaluation methods may overestimate true learning by models, and emphasizes the importance of benchmarking with noise inputs for reliable assessment.

Contribution

It introduces a methodology to distinguish models that genuinely learn from EEG signals versus memorization, highlighting the need for rigorous benchmarking including noise data.

Findings

Model performance on noise data can match EEG data.

Current evaluation practices may overstate model capabilities.

Stricter benchmarking is necessary for reliable EEG-to-Text models.

Abstract

This work critically analyzes existing models for open-vocabulary EEG-to-Text translation. We identify a crucial limitation: previous studies often employed implicit teacher-forcing during evaluation, artificially inflating performance metrics. Additionally, they lacked a critical benchmark - comparing model performance on pure noise inputs. We propose a methodology to differentiate between models that truly learn from EEG signals and those that simply memorize training data. Our analysis reveals that model performance on noise data can be comparable to that on EEG data. These findings highlight the need for stricter evaluation practices in EEG-to-Text research, emphasizing transparent reporting and rigorous benchmarking with noise inputs. This approach will lead to more reliable assessments of model capabilities and pave the way for robust EEG-to-Text communication systems. Code is available at https://github.com/NeuSpeech/EEG-To-Text