Objective
Effective control of neural interfaces is limited by poor signal quality. While neural network-based electroencephalography (EEG) denoising methods for electromyogenic (EMG) artifacts have gained traction in recent years, existing models perform suboptimally in settings with high noise. Since neural interfaces rely almost universally on some form of signal processing, uncovering effective algorithmic insights into EEG denoising is crucial for future success. 

Approach
 To address the shortcomings of current machine learning (ML)-based denoising algorithms, we present a single-channel signal filtration algorithm driven by a new mixture-of-experts (MoE) framework. Our algorithm leverages three new statistical insights into the EEG-EMG denoising problem: (1) EMG artifacts can be partitioned into quantifiable types to aid downstream MoE classification, (2) local experts trained on narrower signal-to-noise ratio (SNR) ranges can achieve performance increases through specialization, and (3) correlation-based objective functions, in conjunction with rescaling algorithms, can enable faster convergence in a neural network-based denoising context.

Main Results
We empirically demonstrate these three insights into EMG artifact removal and use our findings to create a new downstream MoE denoising algorithm consisting of convolutional (CNN) and recurrent (RNN) neural networks. We tested all results on a major benchmark dataset (EEGdenoiseNet) collected from 67 subjects. We found that our MoE denoising model achieved competitive overall performance with contemporary ML denoising algorithms and strong lower bound performance in high noise settings. We further validated the framework on the SEED EEG corpus, observing statistically significant downstream classification gains under high-noise contamination.

Significance
These preliminary results highlight the promise of our MoE framework for enabling advances in EMG artifact removal in EEG, especially in high noise settings. Further experimentation will be necessary to assess our MoE framework on a wider range of test cases and explore its downstream potential to unlock more effective neural interfaces.
&#xD.

Object category learning is a foundational cognitive process. Most human category learning studies involve brief paradigms lasting a few hours and show increased shape tuning in visual areas and task-dependent responses in PFC. Other studies also identify a "frontal bottleneck" that limits multitasking. However, real-world categorization often involves months or years of practice, potentially producing qualitative shifts toward automaticity. We tested the hypothesis that extensive training causes a spatio-temporal shift in the neural circuitry supporting categorization. Participants were trained over >30,000 trials across 5-10 weeks to categorize novel morphed car stimuli via a mobile app. We used fMRI and EEG rapid adaptation techniques to examine neural responses after initial learning (∼4 hr in 1-2 weeks) and after extensive training (∼16 additional hours over another 4-8 weeks). Converging fMRI and EEG results showed that extensive training fundamentally remodeled task-related circuitry: Visual areas in ventral occipito-temporal cortex (vOTC) were initially shape-selective, but category-selective responses emerged in the vOTC after extensive training. The vOTC also showed decreased functional connectivity with the PFC and increased connectivity with motor output areas. This supports the hypothesis that extensive experience enables category decisions to occur outside of the "frontal bottleneck." Critically, the decrease in connectivity between vOTC and PFC was associated with improved categorization performance while dual tasking, indicating increased automaticity. These findings demonstrate that prolonged training reshapes the neural basis of categorization, shifting it from a flexible but attentionally controlled process to a more streamlined, automatic process.

Brain computer interface (BCI) research, as well as increasing portions of the field of neuroscience, have found success deploying large-scale artificial intelligence (AI) pre-training methods in conjunction with vast public repositories of data. This approach of pre-training foundation models using label-free, self-supervised objectives offers the potential to learn robust representations of neurophysiology, potentially addressing longstanding challenges in neural decoding. However, to date, much of this work has focused explicitly on standard BCI benchmarks and tasks, which likely overlooks the multitude of features these powerful methods might learn about brain function as well as other electrophysiological information. We introduce a new method for self-supervised BCI foundation model pre-training for EEG inspired by a transformer-based approach adapted from the HuBERT framework originally developed for speech processing. Our pipeline is specifically focused on low-profile, real-time usage, involving minimally pre-processed data and just eight EEG channels on the scalp. We show that our foundation model learned a representation of EEG that supports standard BCI tasks (P300, motor imagery), but also that this model learns features of neural data related to individual variability, and other salient electrophysiological components (e.g., alpha rhythms). In addition to describing and evaluating a novel approach to pre-training BCI models and neural decoding, this work opens the aperture for what kind of tasks and use-cases might exist for neural data in concert with powerful AI methods.

Current machine learning (ML)-based algorithms for filtering electroencephalography (EEG) time series data face challenges related to cumbersome training times, regularization, and accurate reconstruction. To address these shortcomings, we present an ML filtration algorithm driven by a logistic covariance-targeted adversarial denoising autoencoder (TADA). We hypothesize that the expressivity of a targeted, correlation-driven convolutional autoencoder will enable effective time series filtration while minimizing compute requirements (e.g., runtime, model size). Furthermore, we expect that adversarial training with covariance rescaling will minimize signal degradation. To test this hypothesis, a TADA system prototype was trained and evaluated on the task of removing electromyographic (EMG) noise from EEG data in the EEGdenoiseNet dataset, which includes EMG and EEG data from 67 subjects. The TADA filter surpasses conventional signal filtration algorithms across quantitative metrics (Correlation Coefficient, Temporal RRMSE, Spectral RRMSE), and performs competitively against other deep learning architectures at a reduced model size of less than 400,000 trainable parameters. Further experimentation will be necessary to assess the viability of TADA on a wider range of deployment cases.

Fluid flow dynamics in the brain's ventricles, interstitial spaces, and perivascular spaces, known as the " glymphatic system," are hypothesized to play an important role in brain waste clearance. Healthy function of this complex fluid transporter is most active during sleep, may be critical for maintaining neurological health, and is hypothesized to be important for recovery after acute and chronic injury (e.g. concussion). At present, all sensors for monitoring brain fluid dynamics require invasive contrast agents (e.g. fluorescent dyes injected into cerebrospinal fluid, CSF) and/or are not portable or amenable to long-term repeated monitoring (e.g., magnetic resonance imaging (MRI) methods). We aim to adapt near infrared spectroscopy technologies, which traditionally track hemodynamic activity, to target fluid flow in the glymphatic system and to monitor the temporal dynamics of this water-dominated signal, with an eye toward future applications in continuous portable monitoring. Our goal is to extend frequency domain functional near infrared spectroscopy sensors (FD-fNIRS) to track these CSF-dominated fluid dynamics. In support of this aim, we developed two novel phantoms that mimic key elements of glymphatic system function to demonstrate application of novel FD-fNIRS sensors to human brains in a portable, noninvasive form factor amenable to repeated, continuous testing in a sleep lab-type environment.

The Defense Advanced Research Projects Agency's Revolutionizing Prosthetics program demonstrated the potential for neural interface technologies, enabling patients to control and feel a prosthetic arm and hand, and even pilot an aircraft in simulation. These landmark achievements required invasive, chronically implanted penetrating electrode arrays, which are fundamentally incompatible with applications for the able-bodied warfighter or for long-term clinical applications. Noninvasive neural recording approaches have not been as effective, suffering from severe limitations in temporal and spatial resolution, signal-to-noise ratio, depth penetration, portability, and cost. To help close these gaps, researchers at the Johns Hopkins University Applied Physics Laboratory (APL) are exploring optical techniques that record correlates of neural activity through either hemodynamic signatures or neural tissue motion as represented by the fast optical signal. Although these two signatures differ in terms of spatiotemporal resolution and depth at which the neural activity is recorded, they provide a path to realizing a portable, low-cost, high-performance brain-computer interface. If successful, this work will help usher in a new era of computing at the speed of thought.

The ability to form and store several types of associations between representations of natural images is an area of ongoing research in artificial deep neural networks, which may be informed by biologically-inspired computational models. It is hypothesized that replay of sensory stimuli through cortical-hippocampal connections is responsible for training associations between events, as a powerful form of associative learning. While models of associative memories and sensory processing have been studied extensively, there is a potential for spiking models encompassing sensory processing, reasoning over associations, and learning representations which has not been previously demonstrated. Such networks would be suitable for reasoning and learning from visual data on neuromorphic hardware. In this work, we demonstrate a novel visual reasoning network capable of representing semantic relationships and learning new associations through replay-based association with spiking models using natural images. This is demonstrated through associations of natural images from Tiny Imagenet with a knowledge graph derived from WordNet, and we show that relations in the knowledge graph can be accurately traversed for multiple sequential queries. We also demonstrate learning of a novel association after replayed presentations of natural images. This represents a novel capability for machine learning and reasoning with spiking neural networks which may be amenable to neuromorphic hardware.

Frequency-domain functional near-infrared spectroscopy (FD-fNIRS) has the potential to improve neural imaging of brain hemodynamic responses over conventional magnitude measurements from continuous wave (CW) fNIRS systems by providing additional measurement of phase changes. We evaluated whether phase measurements improved accuracy in decoding motor activity and laterality of movement while recording from motor cortex during a finger tapping task conducted with N=12 subjects. Classification was performed using logistic regression with a single feature derived from hemodynamic response function (HRF) regression. Inspecting the regression results on held-out test data, the majority of subjects showed significant differences between baseline and activity conditions over a typical HRF time course in both magnitude and phase signal components. Combining magnitude and phase information using FD-fNIRS significantly improved classification accuracy of motor conditions at the population level relative to the CW-fNIRS analogue represented by the magnitude signal alone. Our results demonstrate that FD-fNIRS systems can provide benefit over CW-fNIRS for neural decoding applications and are a promising technology for future investigation of non-invasive neural imaging.

We present a 32-transmitter, 32-receiver dual-wavelength frequency-domain (FD) fNIRS system comprised of commercially available avalanche photodiodes, laser drivers and laser mounts. The custom frequency domain (FD) fNIRS system is used to interrogate cerebral tissue with optodes positioned at the posterior occipital region of the head. Data are collected from human subjects watching movie scenes with no sound. We applied cross-validated PCA to identify the number of dimensions retained in the neural signal recorded using FD-fNIRS for the magnitude, phase, and FD (magnitude and phase combined) components of the recorded signal. Importantly, a comparison of the cross-validation error for each signal allows us quantify the dimensionality of the linear subspace spanned by each data type. The number of principal components producing the minimum cross-validation error for the held-out test runs represents the number of orthogonal signal dimensions preserved across training and held-out test data runs. We find that the FD signal captures a higher dimensional space compared to the magnitude or phase signals in isolation. Previous theoretical and empirical work suggest that signals extracted using FD-fNIRS contain higher fidelity neural information than CW-fNIRS in isolation. The findings reported here further support this hypothesis and extend beyond the findings reported in the literature, demonstrating that a higher dimension linear subspace is covered by FD-fNIRS above and beyond the baseline signal captured using traditional CW-fNIRS, assuming other optical performance metrics such as optical dynamic range, noiseequivalent power and cross-talk are comparable. This work was funded by a research contract under Facebook’s Sponsored Academic Research Agreement.

A number of fMRI studies have provided support for the existence of multiple concept representations in areas of the brain such as the anterior temporal lobe (ATL) and inferior parietal lobule (IPL). However, the interaction among different conceptual representations remains unclear. To better understand the dynamics of how the brain extracts meaning from sensory stimuli, we conducted a human high-density electroencephalography (EEG) study in which we first trained participants to associate pseudowords with various animal and tool concepts. After training, multivariate pattern classification of EEG signals in sensor and source space revealed the representation of both animal and tool concepts in the left ATL and tool concepts within the left IPL within 250 ​ms. Finally, we used Granger Causality analyses to show that orthography-selective sensors directly modulated activity in the parietal-tool selective cluster. Together, our results provide evidence for distinct but parallel “perceptual-to-conceptual” feedforward hierarchies in the brain.

Optical neuroimaging technologies aim to observe neural tissue structure and function by detecting changes in optical signals (scatter, absorption, etc…) that accompany a range of anatomical and functional properties of brain tissue. At present, there is a tradeoff between spatial and temporal resolution that is not currently optimized in a single imaging modality. We have developed a coherent optical imaging approach that begins to remove this trade-off and have demonstrated high spatiotemporal (<100µm and >100Hz) in-vivo recordings of neural activity over large 20mm2 areas.

The human visual system can detect objects in streams of rapidly presented images at presentation rates of 70 Hz and beyond. Yet, target detection is often impaired when multiple targets are presented in quick temporal succession. Here, we provide evidence for the hypothesis that such impairments can arise from interference between "top-down" feedback signals and the initial "bottom-up" feedforward processing of the second target. Although it is has been recently shown that feedback signals are important for visual detection, this "crash" in neural processing affected both the detection and categorization of both targets. Moreover, experimentally reducing such interference between the feedforward and feedback portions of the two targets substantially improved participants' performance. The results indicate a key role of top-down re-entrant feedback signals and show how their interference with a successive target's feedforward process determine human behavior. These results are not just relevant for our understanding of how, when, and where capacity limits in the brain's processing abilities can arise, but also have ramifications spanning topics from consciousness to learning and attention.

The grouping of sensory stimuli into categories is fundamental to cognition. Previous research in the visual and auditory systems supports a two-stage processing hierarchy that underlies perceptual categorization: (a) a "bottom-up" perceptual stage in sensory cortices where neurons show selectivity for stimulus features and (b) a "top-down" second stage in higher level cortical areas that categorizes the stimulus-selective input from the first stage. In order to test the hypothesis that the two-stage model applies to the somatosensory system, 14 human participants were trained to categorize vibrotactile stimuli presented to their right forearm. Then, during an fMRI scan, participants actively categorized the stimuli. Representational similarity analysis revealed stimulus selectivity in areas including the left precentral and postcentral gyri, the supramarginal gyrus, and the posterior middle temporal gyrus. Crucially, we identified a single category-selective region in the left ventral precentral gyrus. Furthermore, an estimation of directed functional connectivity delivered evidence for robust top-down connectivity from the second to first stage. These results support the validity of the two-stage model of perceptual categorization for the somatosensory system, suggesting common computational principles and a unified theory of perceptual categorization across the visual, auditory, and somatosensory systems.

Optical neuroimaging technologies aim to observe neural tissue structure and function by detecting changes in optical signals (scatter, absorption, etc…) that accompany a range of anatomical and functional properties of brain tissue. At present, there is a tradeoff between spatial and temporal resolution that is not currently optimized in a single imaging modality. This work focuses on filling the gap between the spatio-temporal resolutions of existing neuroimaging technologies by developing a coherent optics-based imaging system capable of extracting anatomical and functional information across a measurement volume by leveraging a coherent optics-based approach that provides both magnitude and phase information of the sample. We developed a digital holographic imaging (DHI) system capable of detecting these optical signals with a spatial resolution of better than 50 μm over a twenty-five mm2 field of view at sampling rates of 300 Hz and higher. The DHI system operates in the near-infrared (NIR) at 1064 nm, facilitating increased light penetration depths while minimizing contributions from overt changes in oxy- and deoxy-hemoglobin concentration present at shorter NIR wavelengths. This label-free imaging method detects intrinsic signals driven by tissue motion, allowing for innately spatio-temporally registered extraction of anatomical and functional signals in vivo. In this work, we present in vivo results from rat whisker barrel cortex demonstrating signals reflecting anatomical structure and tissue dynamics.

The development of portable non-invasive brain computer interface technologies with higher spatio-temporal resolution has been motivated by the tremendous success seen with implanted devices. This talk will discuss efforts to overcome several major obstacles to viability including approaches that promise to improve spatial and temporal resolution. Optical approaches in particular will be highlighted and the potential benefits of both Blood-Oxygen Level Dependent (BOLD) and Fast Optical Signal (FOS) will be discussed. Early-stage research into the correlations between neural activity and FOS will be explored.

This article examines the current technology- based capabilities of national security and law enforcement officials to assess the credibility of individuals who are being evaluated as a potential source of information or to determine whether they can be trusted with sensitive information. At present, these officials, both domestically and internationally, rely most heavily on the polygraph for a wide variety of credibility assessment applications. However, its accuracy and reliability vary greatly across the different investigative problems to which it is applied. Major improvements in credibility assessment will likely require considerable investments in basic research, but more modest improvements appear within reach by using existing instruments and methods. Perhaps the most promising is the electroencephalogram (EEG), which may be able to detect when an individual is attempting to conceal information. The applicable EEG- based credibility assessment research is reviewed, showing limited but realistic potential for near- term application to some credibility assessment applications.

Pilot-Induced Oscillations (PIOs) are potentially hazardous piloting phenomena in which a pilot's control-inputs and the aircraft control-responses have (for any of a number of possible reasons) become out of phase. During PIOs, aggressive over-controlling on the part of the pilot in order to overcome a perceived lack of control can lead to complete loss of aircraft control. This study shows data recorded from a Cognionics dry electrode system during actual flight exercises can be used on a second-to-second basis to classify whether a pilot was undergoing a PIO event or if a PIO was imminent. If such PIO predictions could be made with adequate accuracy and robustness in real-time, they could form the basis of systems aimed at detecting and/or mitigating PIOs.