Biology subjects

Gilja, V.

Publications and source records attributed to Gilja, V..

11 recordsLinked to original sources

bioRxiv2026

Probabilistic Co-Control in Brain-Computer Interfaces: Uncertainty as a Control Signal in Brain-to-Text Decoding

Neural decoders serve as probabilistic interfaces in co-control brain-to-text BCIs, where predicted uncertainty shapes hypothesis generation and language model integration, enabling decisions to be made safely under uncertainty. However, it remains unclear whether these decoders produce reliable and informative uncertainty, or how training objectives shape these properties. This work characterizes and improves uncertainty representations in brain-to-text decoding. We extend two metrics, calibration error (ECE) and resolution (RES), to evaluate sequential probabilistic predictions from frame-level phoneme estimates to word-level hypotheses, quantifying the reliability and informativeness of model uncertainty. Using this framework, we analyze neural decoders trained with connectionist temporal classification (CTC). To isolate the causal role of uncertainty independent of accuracy, we manipulate predicted probability distributions while holding predicted sequences fixed. Motivated by the observed failures, we further examine the role of the training objective and propose a two-stage cross-entropy (CE) formulation that decouples alignment inference from classification. We show that widely used CTC-trained neural decoders in brain-to-text BCIs produce systematically over-confident predictions, with high confidence persisting even when predictions are incorrect. Controlled manipulations of the prediction reveal that improved ECE and RES enhance hypothesis generation and language-model integration by promoting diverse alternatives and more effective re-ranking of hypotheses aligned with user intent. Mechanistically, CTC relies on over-confident predictions to resolve alignment ambiguity. Replacing CTC with CE loss yields significantly more reliable and informative probabilistic predictions without degrading decoding accuracy. Uncertainty emerges as a system-level design variable in brain-to-text interfaces. Calibrated uncertainty from neural decoders enables effective integration with independently trained language models and reliable error detection. This work reframes uncertainty from a passive output into an active control signal, identifies key components and evaluation criteria for probabilistic co-control, and outlines a pathway toward next-generation BCIs that supports increasingly complex interactions with the world.

bioengineering↗

bioRxiv2026

Compositional decoding of neural activity enhances generalization in handwriting BCIs

Recent brain-computer interfaces (BCIs) have achieved state-of-the-art performance in decoding behavior from neural activity. These models are typically trained on a con-strained set of behaviors, which limits their ability to generalize to real-world settings where behavior is variable, complex, and context-dependent. However, many complex behaviors can be decomposed into a set of reusable behavioral motifs, indicating a compositional organization. Here, we analyze human intracortical neural activity underlying attempted handwriting and find signatures of neural compositionality at a finer resolution than individual letters. We further introduce a compositional temporal decoding model, MOtif-based Temporal Inference Framework (MOTIF), that jointly predicts the fine-scale behavioral motifs (e.g., strokes, phonemes) and the longer-timescale behavior class (e.g., characters, words). We show that the compositional structure leveraged by MOTIF enables improved generalization in few-shot learning. Our results demonstrate that explicitly incorporating compositionality into neural decoders can enhance generalization and sample efficiency, while providing a principled approach to designing more scalable, robust, and interpretable BCIs.

neuroscience↗

bioRxiv2025

SONIC: A Benchmarking Paradigm for Brain-Computer Interfaces

Brain-computer interfaces (BCIs) can restore function for individuals with neuro-logical disorders and have the potential to transform the way people interact with digital systems. However, the development of advanced BCI applications, such as fluent speech synthesis, is dependent on the underlying information transfer capacity of the physical neural interface employed. A significant barrier to progress has been the lack of standardized, application-agnostic methods for benchmarking BCI system performance prior to clinical trials. Here, we introduce SONIC, a novel preclinical benchmarking paradigm designed to evaluate the information transfer rate (ITR) of a BCI system. This paradigm treats the brain and BCI as a noisy communication channel, where information is sent into the brain via precisely controlled sensory stimuli and read out by the neural interface. We implemented this paradigm in an ovine model by presenting rapid sequences of pure tones while recording neural activity from the primary auditory cortex with the Paradromics Connexus(C) BCI, a fully implanted system utilizing high-density intracortical micro-electrode arrays with wireless power and data transmission. A convolutional neural network was used to decode tones based on neural features. Our results demonstrate an achieved ITR of over 200 bits per second (bps), which is the highest reported BCI ITR to date. For reference, this rate exceeds the linguistic information content of human speech. This ITR is achieved with a total neural interface, filtering, and data aggregation delay of 56 milliseconds. Further analysis demonstrated that ITR remains high (> 100 bps) for the lowest total delay tested (11 ms), supporting the needs of latency-sensitive applications (e.g., direct speech synthesis). This work establishes a new benchmark for BCI performance and demonstrates that the Connexus BCI possesses the bandwidth necessary to support highly advanced applications. This benchmark provides a robust framework for preclinical BCI evaluation, enabling principled system design optimization to accelerate the translation of next-generation neurotechnology.

bioengineering↗

bioRxiv2024

Few-shot Algorithms for Consistent Neural Decoding (FALCON) Benchmark

Intracortical brain-computer interfaces (iBCIs) can restore movement and communication abilities to individuals with paralysis by decoding their intended behavior from neural activity recorded with an implanted device. While this activity yields high-performance decoding over short timescales, neural data are often nonstationary, which can lead to decoder failure if not accounted for. To maintain performance, users must frequently recalibrate decoders, which requires the arduous collection of new neural and behavioral data. Aiming to reduce this burden, several approaches have been developed that either limit recalibration data requirements (few-shot approaches) or eliminate explicit recalibration entirely (zero-shot approaches). However, progress is limited by a lack of standardized datasets and comparison metrics, causing methods to be compared in an ad hoc manner. Here we introduce the FALCON benchmark suite (Few-shot Algorithms for COnsistent Neural decoding) to standardize evaluation of iBCI robustness. FALCON curates five datasets of neural and behavioral data that span movement and communication tasks to focus on behaviors of interest to modern-day iBCIs. Each dataset includes calibration data, optional few-shot recalibration data, and private evaluation data. We implement a flexible evaluation platform which only requires user-submitted code to return behavioral predictions on unseen data. We also seed the benchmark by applying baseline methods spanning several classes of possible approaches. FALCON aims to provide rigorous selection criteria for robust iBCI decoders, easing their translation to real-world devices. https://snel-repo.github.io/falcon/

neuroscience↗

bioRxiv2023

Cellular Calcium Activity at Depth Predicted from Surface Potential Recordings using Ultra-high Density Transparent Graphene Arrays

Recording brain activity with high spatial and high temporal resolution across deeper layers of cortex has been a long-sought methodology to study how neural information is coded, stored, and processed by neural circuits and how it leads to cognition and behavior. Electrical and optical neural recording technologies have been the key tools in neurophysiology studies toward a comprehensive understanding of the neural dynamics. The advent of optically transparent neural microelectrodes has facilitated multimodal experiments combining simultaneous electrophysiological recordings from the brain surface with optical imaging and stimulation of neural activity. A remaining challenge is to scale down electrode dimensions to single -cell size and increase the density to record neural activity with high spatial resolution across large areas to capture nonlinear neural dynamics at multiple spatial and temporal scales. Here, we developed microfabrication techniques to create transparent graphene microelectrodes with ultra-small openings and a large, completely transparent recording area. We achieved this by using long graphene microwires without any gold extensions in the field of view. To overcome the quantum capacitance limit of graphene and scale down the microelectrode diameter to 20 m, we used Pt nanoparticles. To prevent open circuit failure due to defects and disconnections in long graphene wires, we employed interlayer doped double layer graphene (id-DLG) and demonstrated cm-scale long transparent graphene wires with microscale width and low resistance. Combining these two advances, we fabricated high-density microelectrode arrays up to 256 channels. We conducted multimodal experiments, combining recordings of cortical potentials with high-density transparent arrays with two-photon calcium imaging from layer 1 (L1) and layer 2/3 (L2/3) of the V1 area of mouse visual cortex. High-density recordings showed that the visual evoked responses are more spatially localized for high-frequency bands, particularly for the multi-unit activity (MUA) band. The MUA power was found to be strongly correlated with the cellular calcium activity. Leveraging this strong correlation, we applied dimensionality reduction techniques and neural networks to demonstrate that single-cell (L2/3) and average (L1 and L2/3) calcium activities can be decoded from surface potentials recorded by high-density transparent graphene arrays. Our high-density transparent graphene electrodes, in combination with multimodal experiments and computational methods, could lead to the development of minimally invasive neural interfaces capable of recording neural activity from deeper layers without requiring depth electrodes that cause damage to the tissue. This could potentially improve brain computer interfaces and enable less invasive treatments for neurological disorders.

neuroscience↗

bioRxiv2022

Imaging through Windansee electrode arrays reveals a small fraction of local neurons following surface MUA

Prior studies have shown that neuronal spikes can be recorded with microelectrode arrays placed on the cortical surface. However, the etiology of these spikes remains unclear. Because the top cortical layer (layer 1) contains very few neuronal cell bodies, it has been proposed that these spikes originate from neurons with cell bodies in layer 2. To address this question, we combined two-photon calcium imaging with electrophysiological recordings from the cortical surface in awake mice using chronically implanted PEDOT:PSS electrode arrays on transparent parylene C substrate. Our electrode arrays (termed Windansee) were integrated with cortical windows offering see-through optical access while also providing measurements of local field potentials (LFP) and multiunit activity (MUA) from the cortical surface. To enable longitudinal data acquisition, we have developed a mechanical solution for installation, connectorization, and protection of Windansee devices aiming for an unhindered access for high numerical aperture microscope objectives and a lifetime of several months while worn by a mouse. Contrary to the common notion, our measurements revealed that only a small fraction of layer 2 neurons from the sampled pool (~13%) faithfully followed MUA recorded from the surface above the imaging field-of-view. Surprised by this result, we turned to computational modeling for an alternative explanation of the MUA signal. Using realistic modeling of neurons with back-propagating dendritic properties, we computed the extracellular action potential at the cortical surface due to firing of local cortical neurons and compared the result to that due to axonal inputs to layer 1. Assuming the literature values for the cell/axon density and firing rates, our modeling results show that surface MUA due to axonal inputs is over an order of magnitude larger than that due to firing of layer 2 pyramidal neurons. Thus, a combination of surface MUA recordings with two-photon calcium imaging can provide complementary information about the input to a cortical column and the local circuit response. Cortical layer I plays an important role in integration of a broad range of cortico-cortical, thalamocortical and neuromodulatory inputs. Therefore, detecting their activity as MUA while combining electrode recording with two-photon imaging using optically transparent surface electrode arrays would facilitate studies of the input/output relationship in cortical circuits, inform computational circuit models, and improve the accuracy of the next generation brain-machine interfaces.

neuroscience↗

bioRxiv2020

The Argo: A 65,536 channel recording system for high density neural recording in vivo

Here we demonstrate the Argo System, a massively parallel neural recording system based on platinum-iridium microwire electrode arrays bonded to a CMOS voltage amplifier array. The Argo system is the highest channel count in vivo neural recording system built to date, supporting simultaneous recording from 65,536 channels, sampled at over 32 kHz and 12-bit resolution. This system is designed for cortical recordings, compatible with both penetrating and surface microelectrodes. We have validated this system by recording spiking activity from 791 neurons in rats and cortical surface Local Field Potential (LFP) activity from over 30,000 channels in sheep. While currently adapted for head-fixed recording, the microwire-CMOS architecture is well suited for clinical translation. Thus, this demonstration helps pave the way for a future high data rate intracortical implant.

neuroscience↗

bioRxiv2020

Local Field Potentials in a Pre-motor Region Predict Learned Vocal Sequences

Neuronal activity within the premotor region HVC is tightly synchronized to, and crucial for, the articulate production of learned song in birds. Characterizations of this neural activity typically focuses on patterns of sequential bursting in small carefully identified subsets of single neurons in the HVC population. Much less is known about population dynamics beyond the scale of individual neurons. There is a rich history of using local field potentials (LFP), to extract information about behavior that extends beyond the contribution of individual cells. These signals have the advantage of being stable over longer periods of time and have been used to study and decode complex motor behaviors, such as human speech. Here we characterize LFP signals in the putative HVC of freely behaving male zebra finches during song production, to determine if population activity may yield similar insights into the mechanisms underlying complex motor-vocal behavior. Following an initial observation that structured changes in the LFP were distinct to all vocalizations during song, we show that it is possible to extract time varying features from multiple frequency bands to decode the identity of specific vocalization elements (syllables) and to predict their temporal onsets within the motif. This demonstrates that LFP is a useful signal for studying motor control in songbirds. Surprisingly, the time frequency structure of putative HVC LFP is qualitatively similar to well established oscillations found in both human and non-human mammalian motor areas. This physiological similarity, despite distinct anatomical structures, may give insight to common computational principles for learning and/or generating complex motor-vocal behaviors. Author SummaryVocalizations, such as speech and song, are a motor process that requires the coordination of several muscle groups receiving instructions from specific brain regions. In songbirds, HVC is a premotor brain region required for singing and it is populated by a set of neurons that fire sparsely during song. How HVC enables song generation is not well understood. Here we describe network activity in putative HVC that precedes the initiation of each vocal element during singing. This network activity can be used to predict both the identity of each vocal element (syllable) and when it will occur during song. In addition, this network activity is similar to activity that has been documented in human, non-human primate, and mammalian premotor regions tied to muscle movements. These similarities add to a growing body of literature that finds parallels between songbirds and humans in respect to the motor control of vocal organs. Given the similarities of the songbird and human motor-vocal systems these results suggest that the songbird model could be leveraged to accelerate the development of clinically translatable speech prosthesis.

neuroscience↗

bioRxiv2019

Microscale physiological events on the human cortical surface detected with PEDOT:PSS Electrodes

Despite ongoing advancements in our understanding of the local single-cellular and network-level activity of neuronal populations in the human brain, extraordinarily little is known about their intermediate microscale local circuit dynamics. Here, we utilized ultrahigh density microelectrode arrays and a rare opportunity to perform intracranial recordings across multiple cortical areas in human participants to discover three distinct classes of cortical activity that are not locked to ongoing natural brain rhythmic activity. The first included fast waveforms similar to extracellular single unit activity. The other two types were discrete events with slower waveform dynamics and were found preferentially in upper layers of the grey matter. They were also observed in rodents, non-human primates, and semi-chronic recordings in humans via laminar and Utah array microelectrodes. The rates of all three events were selectively modulated by auditory and electrical stimuli, pharmacological manipulation, and cold saline application and had small causal co-occurrences. These results suggest that with the proper combination of high resolution microelectrodes and analytic techniques it is possible to capture neuronal dynamics that lay between somatic action potentials and aggregate population activity and that understanding these intermediate microscale dynamics may reveal important details of the full circuit behavior in human cognition.

neuroscience↗

bioRxiv2017

Sequence Transfer Learning for Neural Decoding

A fundamental challenge in designing brain-computer interfaces (BCIs) is decoding behavior from time-varying neural oscillations. in typical applications, decoders are constructed for individual subjects and with limited data leading to restrictions on the types of models that can be utilized. currently, the best performing decoders are typically linear models capable of utilizing rigid timing constraints with limited training data. Here we demonstrate the use of Long Short-Term Memory (LSTM) networks to take advantage of the temporal information present in sequential neural data collected from subjects implanted with electrocorticographic (ECoG) electrode arrays performing a finger flexion task. our constructed models are capable of achieving accuracies that are comparable to existing techniques while also being robust to variation in sample data size. Moreover, we utilize the LSTM networks and an affine transformation layer to construct a novel architecture for transfer learning. We demonstrate that in scenarios where only the affine transform is learned for a new subject, it is possible to achieve results comparable to existing state-of-the-art techniques. The notable advantage is the increased stability of the model during training on novel subjects. Relaxing the constraint of only training the affine transformation, we establish our model as capable of exceeding performance of current models across all training data sizes. Overall, this work demonstrates that LSTMS are a versatile model that can accurately capture temporal patterns in neural data and can provide a foundation for transfer learning in neural decoding.

neuroscience↗

bioRxiv2017

A neural decoder for learned vocal behavior

Brain Machine Interfaces (BMIs) hold promise to restore impaired motor function and, because they decode neural signals to infer behavior, can serve as powerful tools to understand the neural mechanisms of motor control. Yet complex behaviors, such as vocal communication, exceed state-of-the-art decoding technologies which are currently restricted to comparatively simple motor actions. Here we present a BMI for birdsong, that decodes a complex, learned vocal behavior directly from neural activity.

neuroscience↗