A new deep learning system can tell, just from your brain's electrical signals, whether you're resting or actively doing math, memorizing something, or listening to music.
Researchers combined a convolutional neural network (CNN) with a gated recurrent unit (GRU), two AI techniques suited to finding patterns in sequences, to classify EEG brain-wave recordings. Instead of feeding the model raw signals, they extracted time-frequency features first, then tested the setup against several other deep learning and machine learning classifiers, including a proposed architecture called 2D-Net. The combined approach won each comparison. On three separate rest-versus-task tests, it hit 83.177% accuracy for math, 76.107% for memory, and 83.432% for music.
EEG is cheap and non-invasive compared to other brain-imaging methods, which is why it keeps showing up in brain-computer interface and attention-monitoring research. The real finding here isn't the headline accuracy number, it's the spread: memory tasks were nearly seven points harder to detect than math or music. That gap suggests some mental states leave a much fainter electrical footprint than others, a problem no amount of model tuning fully erases.
Read this as a lab benchmark, not a mind-reading device - a plausible research architecture that still needs work before anyone builds a product on it.