91̽»¨

Accessing information in learned representations is critical for scientific discovery in high-dimensional domains. We introduce a novel method based on latent flow matching with classifier-free guidance that disentangles latent subspaces by explicitly separating information included in conditioning from information that remains in the residual representation. Across three experiments -- a synthetic 2D Gaussian toy problem, colored MNIST, and the Galaxy10 astronomy dataset -- we show that our method enables access to meaningful features of high dimensional data. Our results highlight a simple yet powerful mechanism for analyzing, controlling, and repurposing latent representations, providing a pathway toward using generative models for scientific exploration of what we don't capture, consider, or catalog.

Introduction: The normal transition to sleep is characterized by a reduction in higher frequency activity and an increase in lower frequency activity in frontal brain regions. In sleep onset insomnia these changes in activity are weaker and may prolong the transition to sleep.

Methods: Using a wearable device, we compared 30min of short duration repetitive transcranial electric stimulation (SDR-tES) at 0.75Hz, prior to going to bed, with an active control at 25Hz in the same individuals.

Results: Treatment with 0.75Hz significantly reduced sleep onset latency (SOL) by 53% when compared with pre-treatment baselines and was also significantly more effective than stimulation with 25Hz which reduced SOL by 30%. Reductions in SOL with 25Hz stimulation displayed order effects suggesting the possibility of placebo. No order effects were observed with 0.75Hz stimulation. The decrease in SOL with 0.75Hz treatment was proportional to an individual’s baseline wherein those suffering from the longest pre-treated SOLs realized the greatest benefits. Changes in SOL were correlated with left/right frontal EEG signal coherence around the stimulation frequency, providing a possible mechanism and target for more focused treatment. Stimulation at both frequencies also decreased perceptions of insomnia symptoms measured with the Insomnia Severity Index, and comorbid anxiety measured with the State Trait Anxiety Index.

Discussion: Our study identifies a new potential treatment for sleep onset insomnia that is comparably effective to current state-of-practice options including pharmacotherapy and cognitive behavioral therapy and is safe, effective, and can be delivered in the home.

The progression of human-computer interfaces into immersive and touchless realities requires new ways of interacting with machines that are correspondingly intuitive and seamless. Among these are gesture-based systems that use natural hand movements to interact with and control digital devices. Today, these systems are most commonly implemented through the use of cameras or inertial sensors, which have drawbacks in environments that are poorly lit, in conditions where the hands are obscured, or for applications that require fine motor control. More recent studies have advocated for the use of surface electromyography (sEMG) to capture gesture information by sensing electrical activity generated by muscle contraction. While promising demonstrations have been shown, studies have also outlined limitations in sEMG when it comes to generalization across a population, largely due to physiological differences between individuals. Magnetomyography (MMG) is an alternative modality for measuring the same motor signals at the muscle, but is impervious to distortions caused by tissue, hair, and moisture; this indicates potential for lower variability caused by physiological differences and changes in skin conductivity, making MMG a promising generalizable solution for gesture control. To test this theory, we developed wristbands with magnetic sensors and implemented a signal processing pipeline for gesture classification. Using this system, we measured MMG across 30 participants performing a gesture task consisting of nine discrete gestures. We demonstrate average single-participant classification accuracy of 95.4%, rivaling state-of-the-art accuracy with sEMG. In addition, we achieved higher cross-session and cross-participant accuracy compared to sEMG studies. Given that these results were obtained with a non-ideal recording system, we anticipate significantly better results with better sensors. Together, these findings suggest that MMG can provide higher performance for control systems based on gesture recognition by overcoming limitations of existing techniques.