Spatiotemporal Convolutions on EEG signal -- A Representation Learning Perspective on Efficient and Explainable EEG Classification with Convolutional Neural Nets

The paper's real contribution isn't just comparing architectures, but demonstrating that mathematically equivalent formulations can produce different learned representations and generalization behavior. This surfaces a gap between theoretical equivalence and empirical learning dynamics that practitioners often overlook when choosing between 1D and 2D convolutions.

This aligns directly with the physics-informed GNN work from May 5th, which showed that explicitly encoding domain constraints into computation graphs improves both efficiency and interpretability. Both papers reject the assumption that neural networks will implicitly discover the right inductive bias. For BCI specifically, the temporal-spatial coupling in EEG mirrors the gauge-invariance problem in lattice systems: forcing the architecture to respect the domain structure upfront beats hoping backprop finds it. The clinical readmission prediction paper from May 1st also touches this tension, though in a different context (observation window selection), showing that architectural choices around temporal encoding consistently matter more than practitioners initially expect.

If this work leads to published benchmarks on standard BCI datasets (like BCI Competition IV) showing consistent accuracy gains over 1D baselines in both offline and online (closed-loop) settings, that confirms the finding generalizes beyond the motor imagery task tested here. If gains disappear or flip sign when tested on different subject populations or electrode montages, that signals the benefit is dataset-specific rather than fundamental to EEG processing.