Overview
The One-Block Transformer (1BT) is a proposed architecture designed for the expeditious and compact assessment of cognitive workload using electroencephalography (EEG) data. This model aims to balance representational capacity with computational efficiency, addressing challenges in practical deployment for continuous cognitive workload estimation.
Research Context
Accurate and continuous estimation of cognitive workload is portrayed as fundamental for the development of adaptive human-machine systems. The existing challenge in this domain concerns the design of architectures that maintain a balance between their representational capacity and computational efficiency, particularly for applications requiring practical deployment.
Approach
The 1BT model aggregates multi-channel temporal sequences through a minimal latent bottleneck. This process involves the utilization of a single cross-attention module, subsequently followed by a lightweight self-attention mechanism. To evaluate the model, a controlled study was conducted involving 11 participants. These participants engaged in three distinct cognitively diverse tasks: abstract reasoning, numerical problem-solving, and an interactive video game. During these tasks, continuous EEG recordings were collected across two specified workload levels. A systematic architectural analysis was performed to identify the most compact configuration capable of preserving high performance while simultaneously reducing computational cost.
Findings
- The 1BT model achieved high workload classification performance.
- The final model operates with fewer than 0.5 million parameters.
- The final model demonstrated a computational cost of 0.02 GFLOPs.
- This performance and efficiency were maintained while substantially lowering computational cost compared to other configurations.
Why This Matters
The proposed design direction, characterized by the compact nature and efficiency of the 1BT model, is indicated as suitable for real-time cognitive workload monitoring within resource-constrained settings.