A lightweight architecture for non-invasive brain-to-text synthesis using sparse semantic anchors and Large Language Models
Year: 2026
Type: dissertation
Abstract: The transformation of non-invasive electroencephalography (EEG) signals into natural language is a frontier in Brain-Computer Interfaces (BCIs), offering transformative potential for assistive communication. However, existing paradigms predominantly rely on end-to-end fine-tuning of Large Language Models (LLMs) on raw neural data, incurring significant computational costs, risking neural data privacy, and suffering from "subject-gap" hallucinations. This thesis presents a two-phase research trajectory toward efficient, privacy-preserving neural decoding. In the first phase, we analyze the representational capabilities of continuous neural embeddings, revealing that diffusion-based latent spaces are highly discriminative of subject identity rather than semantic content, leading to severe model hallucinations in zero-shot contexts. Addressing these bottlenecks, the second phase introduces SENSE (SEmantic Neural Sparse Extraction), a modular framework that decouples decoding into local semantic retrieval and prompt-based language generation. By projecting EEG signals into a discrete, CLIP-aligned semantic space, SENSE extracts human-interpretable "Bag-of-Words" (BoW) anchors. These anchors serve as structured prompts for off-the-shelf LLMs to synthesize fluent text without parameter adaptation. Our evaluations demonstrate that SENSE, with only 6 million parameters, achieves competitive performance in open-vocabulary decoding while ensuring that sensitive raw neural signals remain on-device. This research establishes a lightweight, scalable, and interpretable pathway for the next generation of privacy-aware neural communication systems.
DOI: 10.26153/tsw/64325
Author: Murhekar, Akshaj
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7171532740