Research map / Architectures and efficiency / RNNs and normalization
Neural Bookmarks: Information Retrieval with Deep Learning and EEG Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes neural information retrieval, where a document is retrieved by mentally recalling it with EEG, adapting a prior EEG memory decoder (EMD) and testing its data size, model size, and channel count for practicality.
Problem
Finding previously seen documents on the internet is hard, and keyword bookmarks are subjective and costly to create; the paper asks whether EEG-based memory decoding can practically support document retrieval by recall.
Approach
Defines a neural information retrieval application in which users index documents by recording EEG while recalling them and later retrieve them by recalling again; adapts the EMD segment-embedding plus KNN ensemble design with precomputed indexing and per-document training across days, and evaluates reduced data, smaller models, and fewer EEG channels.
Key results
- Full-data EMD with 655k parameters and 14 channels reaches mean top-1 0.750 and top-3 0.944 on 25-topic test sets.
- Small-data regime (40 training documents, 8 s indexing/search EEG) drops top-1 to 0.476 and top-3 to 0.812.
- Reducing to 39k parameters and 4 channels with small data gives top-1 0.500 and top-3 0.852, with 4 channels matching 14 channels at full data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| reduce training documents from 78 to 40 and EEG recording duration (ablation) | EMD · EMD test topics | top-1 accuracy: 0.750 [0.71, 0.78] -> 0.476 [0.44, 0.51] | Hurt |
| reduce training documents from 78 to 40 and EEG recording duration (ablation) | EMD · EMD test topics | top-3 accuracy: 0.944 [0.93, 0.96] -> 0.812 [0.78, 0.84] | Hurt |
| reduce embedding model from 655k to 39k parameters (ablation) | EMD · EMD test topics | top-1 accuracy: 0.750 [0.71, 0.78] -> 0.690 [0.66, 0.72] | Hurt |
| reduce embedding model from 655k to 39k parameters (ablation) | EMD · EMD test topics | top-3 accuracy: 0.944 [0.93, 0.96] -> 0.918 [0.90, 0.94] | Hurt |
| reduce EEG channels from 14 to 4 (ablation) | EMD · EMD test topics | top-1 accuracy: 0.750 [0.71, 0.78] -> 0.764 [0.73, 0.80] | No clear effect |
| reduce EEG channels from 14 to 4 (ablation) | EMD · EMD test topics | top-3 accuracy: 0.944 [0.93, 0.96] -> 0.954 [0.93, 0.97] | No clear effect |
| reduce training data, embedding model to 39k parameters, and EEG channels to 4 (ablation) | EMD · EMD test topics | top-1 accuracy: 0.750 [0.71, 0.78] -> 0.500 [0.46, 0.55] | Hurt |
| reduce training data, embedding model to 39k parameters, and EEG channels to 4 (ablation) | EMD · EMD test topics | top-3 accuracy: 0.944 [0.93, 0.96] -> 0.852 [0.82, 0.89] | Hurt |
Surprising findings
- Using only four prefrontal and frontal electrodes produced no significant change in top-1 or top-3 accuracy relative to 14 channels.
- Reducing training data had little impact once the embedding model was reduced in size.
Limitations
- Accuracy evaluated on only 25-topic test sets; scaling to thousands of documents untested.
- Effect of noise in consumer recording environments and person-to-person variation not assessed.
- Retrieval after long time gaps (weeks to years) not evaluated.
Experimental setup
Datasets: EEG recordings of recalled Wikipedia pages (103 pages, 16-channel EEG, from Bruns, Haidar, and Rubino 2023)
Benchmarks: in-house EEG Wikipedia page recall dataset
Metrics: top-1 accuracy top-3 accuracy
Baselines: EMD (full data, 655k parameters, 14 channels)
Builds on: Bruns, Haidar, and Rubino 2023 (EMD) supervised contrastive loss k-nearest neighbor
Closest papers by meaning
- Decoding Natural Images from EEG for Object Recognition (ICLR 2024 · similarity 0.83)
- HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (NeurIPS 2024 · similarity 0.82)
- Neural decoding from stereotactic EEG: accounting for electrode variability across subjects (NeurIPS 2024 · similarity 0.81)
- CognitionCapturer: Decoding Visual Stimuli from Human EEG Signal with Multimodal Information (AAAI 2025 · similarity 0.81)
- Brain decoding: toward real-time reconstruction of visual perception (ICLR 2024 · similarity 0.81)
- NeurIPT: Foundation Model for Neural Interfaces (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.