Brain decoding gets cheaper: behavior-only training and foundation models that skip fine-tuning

Decoding without the neural data

The bottleneck in neural motor decoding has always been data: paired neural and behavioral recordings are expensive and invasive to collect. A preprint this week proposes BEEMO, a framework that decouples neural and behavioral representation learning, so training can use arbitrary mixtures of paired recordings and unpaired behavioral data. The headline result is that behavior-only pretraining achieved performance comparable to models pretrained on substantially larger neural datasets, and incorporating unpaired behavioral data improved decoding performance, particularly when transferring to new tasks. The practical claim is that abundant behavioral trajectories, which are cheap, can substitute for much of the neural data that is not. m:…9639

That is a scaling argument with a clinical edge. If behavioral data can carry the pretraining load, neural decoding models can generalize across subjects and tasks without requiring massive paired datasets for each new one. Incorporating unpaired behavioral data improves decoding performance, particularly when transferring to new tasks. But the direction is clear, and it mirrors what happened in language models, where cheap abundant data replaced expensive curated data as the scaling fuel. m:…9639

Foundation models for EEG

A second preprint brings the foundation-model playbook to EEG. Researchers probed the attention maps of the REVE EEG foundation model to generate functional connectivity matrices, then used lightweight linear probing for emotion recognition. In a preprint, the approach required two orders of magnitude fewer calculations than standard encoder fine-tuning, while matching the performance of fully fine-tuned architectures. The attention-derived connectivity edges also showed a significant over-representation of frontal region connections, consistent with prior affective-computing literature, which gives the method an interpretability story as well as an efficiency one. m:…6705

Attention-derived functional connectivity matrices from the REVE foundation model are sufficient for emotion recognition without fine-tuning. The point is deployment. Foundation models are only useful in brain-computer interfaces if they can run in resource-constrained settings, and fine-tuning a full encoder for every application is the opposite of that. Linear probing over attention-derived features is a way to get foundation-model performance at a fraction of the cost, and the authors frame it as a path toward real-time, interpretable neurotechnology. m:…6705

Sleep staging and the concussion signal

A third preprint shows the same stack working in real time. A causal tracker using a 10-second EEG window can estimate N2 and N3 sleep stages every 2 seconds, with an AUC of 0.967 when scalp and thalamic signals are combined. Thalamic channels alone outperformed mesial temporal channels, which makes thalamic leads a viable non-scalp alternative for triggering stimulation. m:…6706

A fourth preprint, on concussion, shows what these signals can measure. Researchers analyzed longitudinal neuroimaging data from collegiate athletes and found BOLD signal variability decreased in somatomotor and language networks within 48 hours of injury, correlating with self-reported symptoms and computerized cognitive performance. The measure captures individual differences in neural, vascular, and neurovascular coupling that traditional mean activation measures fail to capture. An updated Human Neocortical Neurosolver model, with human-specific dendritic and synaptic dynamics, gives researchers a biophysical tool for interpreting these non-invasive signals, simulating dendritic calcium spikes in layer 5 neurons and shorter dendritic spikes in layer 2/3 pyramidal neurons. Its updated circuitry allows researchers to disentangle somatic, synaptic, and dendritic contributions to human EEG and MEG signatures. m:…6707 m:…7978