Alljoined-1.6M: A Million-Trial EEG-Image Dataset for Evaluating Affordable Brain-Computer Interfaces

TL;DR

This paper introduces a large-scale, affordable EEG dataset with over 1.6 million trials, demonstrating that high-level semantic decoding and EEG-to-Image reconstruction are feasible using consumer-grade hardware, thus enabling accessible BCI research.

Contribution

The creation of a large, open-source EEG dataset using low-cost hardware, enabling scalable and cost-effective brain-computer interface research and semantic decoding.

Findings

Semantic information can be decoded from consumer-grade EEG data.

EEG-to-Image reconstruction is effective despite lower signal quality.

Decoding performance improves log-linearly with data volume.

Abstract

We present a new large-scale electroencephalography (EEG) dataset as part of the THINGS initiative, comprising over 1.6 million visual stimulus trials collected from 20 participants, and totaling more than twice the size of the most popular current benchmark dataset, THINGS-EEG2. Crucially, our data was recorded using a 32-channel consumer-grade wet electrode system costing ~$2.2k, around 27x cheaper than research-grade EEG systems typically used in cognitive neuroscience labs. Our work is one of the first open-source, large-scale EEG resource designed to closely reflect the quality of hardware that is practical to deploy in real-world, downstream applications of brain-computer interfaces (BCIs). We aim to explore the specific question of whether deep neural network-based BCI research and semantic decoding methods can be effectively conducted with such affordable systems, filling an important gap in current literature that is extremely relevant for future research. In our analysis, we not only demonstrate that decoding of high-level semantic information from EEG of visualized images is possible at consumer-grade hardware, but also that our data can facilitate effective EEG-to-Image reconstruction even despite significantly lower signal-to-noise ratios. In addition to traditional benchmarks, we also conduct analyses of EEG-to-Image models that demonstrate log-linear decoding performance with increasing data volume on our data, and discuss the trade-offs between hardware cost, signal fidelity, and the scale of data collection efforts in increasing the size and utility of currently available datasets. Our contributions aim to pave the way for large-scale, cost-effective EEG research with widely accessible equipment, and position our dataset as a unique resource for the democratization and development of effective deep neural models of visual cognition.