Challenge 2026
Advancing Brain-Computer Interfaces with Neuromorphic Systems: evaluating spike-based motor imagery decoders for absolute classification accuracy and compute efficiency.
Inaugural Neuromorphic Benchmark Competition
The THOR Neuromorphic Commons partnered with NeuroBench and the Edge AI Foundation to benchmark spiking neural network architectures on continuous EEG Motor Imagery (OpenBMI). Submissions were evaluated in two distinct tracks: maximum classification accuracy and energy-efficient synaptic operations.
Dataset
A Brain-Computer Interface (BCI) translates neural activity into commands for external software or hardware. This challenge specifically targets the Motor Imagery (MI) paradigm.
Moving neural decoding technology out of bulky clinical setups into ultra-low-power wearable decoders for motorized wheelchairs and robotic prosthetics.
Life-changing assistive potential for individuals living with stroke recovery, ALS, or spinal cord injuries.
The OpenBMI dataset provides an unprecedented scale of high-quality EEG recordings, distinguishing it from smaller legacy datasets that suffer from severe overfitting.
Reference Publication
Lee, Min-Ho, et al. “EEG dataset and OpenBMI toolbox for three BCI paradigms: An investigation into BCI illiteracy.” GigaScience 8.5 (2019): giz002.
Read paper in GigaScience
Challenge Overview
Standard deep neural networks ingest continuous, floating-point analog signals. Neuromorphic systems, however, operate on event-driven, asynchronous binary spikes. Models must convert continuous amplitudes into discrete spike trains to classify left-hand vs. right-hand motor imagery.
Measures the baseline capability of the neuromorphic model to accurately classify left-hand versus right-hand motor imagery across the cross-validation splits, independent of compute cost or model footprint.
Rather than dense Multiply-Accumulates (MACs), the NeuroBench harness measures effective Synaptic Operations (SynOps = MACs + 0.1 · ACs) triggered by sparse spike trains, paired with memory footprint (weights & state).
Instructions
Official repositories, datasets, instructions, and submission documentation provided to all challenge participants.
Sign up to participate in the THOR NeuroBench Challenge 2026.
Detailed guidelines and rules for the challenge.
Instructions for the NeuroBench Code Harness with examples.
Preprocessed Dataset for OpenBMI Challenge : EEG-MI Task.
Upload your model code securely.
Timeline
Milestones from official announcement to Phase-I leaderboard release and future Phase-II invitations.
Official launch of Phase-I of the THOR x Neurobench Challenge, for developing neuromorphic spiking algorithms.
Release of the openBMI EEG-MI dataset and detailed specifications for the Phase-I challenge, focusing on efficient neuromorphic algorithm development.
Registration opens for participants interested in tackling the BCI challenge with neuromorphic spiking algorithms.
The portal for submitting neuromorphic spiking algorithm solutions for Phase-I of the challenge becomes active.
Final deadline for all Phase-I submissions of neuromorphic spiking algorithms for the EEG-MI task.
Announcement of top-performing neuromorphic algorithms from Phase-I
Extending invitations for the advanced Phase-II.
The second phase of the challenge, focusing on further refinement and deployment of neuromorphic spiking algorithms for BCI applications.
Leaderboard
Final competition standings across Track 1 (Absolute Accuracy) and Track 2 (Compute Efficiency). All submissions evaluated on the unified NeuroBench EEG-MI test harness.
Ranked purely on the highest reported test accuracy, independent of model size, compute, or memory footprint. A team's result is its single best submission. There is no penalty for the number of attempts.
Each point represents one submission, plotted by date against verified test accuracy.
Teams ranked by top submission. Click any expandable row to inspect that team's iterative submissions.
Ranks the trade-off between accuracy, compute (effective MACs + effective ACs), and memory footprint. Two independent scoring strategies were developed and are both reported below for complete transparency.
NES = Accuracy − λ · log₁₀(R), where R = √[(W / OPS_REF) · (Footprint / FP_REF)]
Accuracy minus a log-penalty for compute + footprint, self-normalized per submission without pool distortion.
Mathematical Formulation
Log-scaled x-axis. Ringed points sit on the non-dominated Pareto front.
Each team's single highest-scoring submission under Strategy 1.
Click any team with multiple submissions to expand full submission records.
Final Score = Accuracy × Model Efficiency, where Model Efficiency = Ops Savings^0.5 · Footprint Savings^0.5
Accuracy × Savings Score, normalized against the pool, but mathematically sensitive to pool outliers.
Mathematical Formulation
Log-scaled x-axis showing memory footprint in bytes vs. test accuracy.
Each team's highest-scoring submission under Strategy 2.
Click any team to expand all submission records.
Awards
Recognizing breakthroughs in brain-computer interfaces with neuromorphic computing, generously sponsored by the Edge AI Foundation.
$500
Awarded to the top-ranking submission for absolute classification accuracy on the OpenBMI EEG-MI benchmark test splits.
$500
Awarded to the top-performing submission optimizing the joint Pareto frontier of high classification accuracy, sparse synaptic operations, and tiny memory footprint.
People & Partners
The steering committee and partner organizations behind the THOR x NeuroBench Challenge 2026.
Collaborating Entities
Sponsoring Organizations
