CRE-TSCAE: A Novel Classification Model Based on Stacked Convolutional Autoencoder for Dual-Target RSVP-BCI Tasks

Area of Science:

• Neuroscience
• Machine Learning
• Biomedical Engineering

Background:

• Rapid Serial Visual Presentation (RSVP) is crucial for target identification in surveillance but primarily studied for single targets.
• Dual-target RSVP-BCI (Brain-Computer Interface) research is limited, hindering broader applications.

Purpose of the Study:

• To develop an effective classification model for detecting two targets and one non-target in RSVP tasks.
• To address the limitations of existing RSVP-BCI algorithms in dual-target scenarios.

Main Methods:

• Proposed a novel Common Representation Extraction-Targeted Stacked Convolutional Autoencoder (CRE-TSCAE) model.
• CRE extracts common representations to reduce intra-class variability and enhance inter-class distinction.
• TSCAE controls training uncertainty and minimizes the need for extensive target training data.

Main Results:

• CRE-TSCAE validated on World Robot Contest 2021 and 2022 ERP datasets.
• Achieved an average accuracy (ACC) of 71.25%, outperforming state-of-the-art RSVP decoding algorithms by at least 6.5%.
• Demonstrated strong ability in extracting discriminative latent features for differentiating targets.

Conclusions:

• CRE-TSCAE offers an innovative and effective solution for dual-target RSVP-BCI.
• Provides insights into neurophysiological distinctions between different targets.
• Enhances classification accuracy in complex RSVP tasks.