GRU models outperform Mamba for brain-to-text decoding

Researchers have investigated the effectiveness of different model architectures and target representations for intracortical brain-to-text systems. The study compared gated recurrent units (GRUs) with selective state-space models (Mamba) and evaluated both phonetic and character-level targets on the Brain-to-Text '25 benchmark. Results indicated that GRUs performed best, with phonetic targets achieving 12.62% PER and character targets reaching 13.39% CER after language model rescoring. The Mamba hybrid models were competitive but did not outperform the GRU baselines. AI

IMPACT This research contributes to the understanding of optimal model architectures and data representations for brain-computer interfaces, potentially improving future assistive technologies.

RANK_REASON Academic paper detailing novel research findings on model architectures and target representations for a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• Brain-to-Text '25
• gated recurrent unit
• Jose A. Gonzalez-Lopez
• Mamba

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →