NeuroSketch is a new recipe for building neural networks that decode brain signals, and it just beat ten rival designs across the board.
Researchers started by testing nine basic architectures against each other on neural decoding tasks and found that CNN-2D, which treats brain signals as a two-dimensional grid, won out. They then refined it with macro-level tweaks (gradually expanding feature maps, downsampling early) and micro-level ones (grouped convolutions) to keep the model efficient. The result is two variants, NeuroSketch-Base at 1.4 million parameters and NeuroSketch-Large at 4.2 million, validated across nearly 5,000 experiments on eight tasks covering visual, auditory, and speech signals recorded via EEG, SEEG, and ECoG. Against ten baseline models, the two versions together posted the best accuracy on every task.
Most brain-computer interface work treats architecture as a formality, borrowing whatever design worked for images or text and hoping it transfers. NeuroSketch instead tests that assumption directly, offering evidence for which structural choices help versus which are just inherited habit. That kind of groundwork matters in a field increasingly aimed at real medical devices, not just lab demos.
The code is open-source, so the real verdict comes when other labs run it on their own data rather than the group's own five thousand experiments.