During the MI task, subjects were prompted with visual cues to imagine the kinetic movement of either their left hand or right hand (a two-class classification problem). Each trial consists of a 4-second active imagination phase, flanked by resting baselines. This yields 100 trials per class, per session, for each individual subject. Participants must design a neuromorphic model architecture to classify the EEG-MI signals as either left or right hand movements.

Standard Deep Neural Networks (DNNs) ingest continuous, floating-point analog signals. However, neuromorphic architectures operate on event-driven, binary logic. Therefore, the continuous EEG amplitudes must be converted into discrete, asynchronous spike trains before processing. For this challenge, the provided data must be formatted as sparse, multi-dimensional binary tensors (representing 1 for a spike, 0 for no spike). Participants are encouraged to explore varying spike encoding schemes to optimize their models for Track 1, but have to adhere to a standardized (provided) spike format for Track 2.

To rigorously evaluate the real-world viability of the submitted neuromorphic models, the challenge implements a strict, multi-phase cross-validation strategy designed to test generalization capabilities.

Track 1 : Classification Accuracy

The baseline capability of the model to correctly classify left-hand versus right-hand motor imagery. A baseline benchmark of average accuracy across the test splits will be provided to all participants after registration.

Track 2 : Compute Efficiency

Instead of counting dense Multiply-Accumulate (MAC) operations used in standard GPUs, the harness calculates the number of effective Synaptic Operations (SynOps) triggered by the sparse spike trains. Fewer spikes and sparser activations mean lower energy consumption on the THOR architecture. In addition, submissions in Track 2 will be evaluated based on their memory usage (Footprint).