CRCNS TALKS
AUDITION
MOLECULAR / CELLULAR LEVEL
SPATIAL NAVIGATION
VISION
SENSORI-MOTOR / DECISION MAKING
MODELING / THEORY
SYSTEMS
AUDITION
Laminar Processing in Auditory Cortex ‘Untangles’ Sound Representations During Active Listening
Srivatsun Sadagopan, University of Pittsburgh School of Medicine
Abstract:Vocalizations, such as animal calls and human speech, are produced with tremendous between-subject and inter-trial variability. A central function of auditory processing is to generalize over this variability and group calls into discrete categories. Previously, we developed an interpretable hierarchical model that accomplishes categorization by detecting features of intermediate complexity that capture the ‘gist’ of each call category. Our prior experimental results also broadly validated this model. Specifically, neural responses selective for intermediate features emerged in the superficial layers (L2/3) of primary auditory cortex (A1), and behavioral choices in a call categorization task were well-explained by the feature-based model. Here, we ask how call representations in different A1 laminae are modulated by task performance. We performed neural recordings, using chronically implanted Neuropixels probes, from A1 of guinea pigs trained to categorize conspecific vocalizations. Feature-selective neurons in A1 L2/3, while preserving their high selectivity to specific call features, showed increased output gain during active task performance. Incorporating this increased output gain into our model resulted in better separated (‘untangled’) representations of categories. Together, these theoretical and experimental results reveal novel computational principles underlying auditory categorical representations and their modulation by attention. More broadly, our studies may provide insight into general computational principles underlying categorization across sensory modalities.
Encoding Auditory Space in Marmoset Cortex
Yi Zhou, Arizona State University
Abstract: A century ago Jakob von Uexküll argued that every species inhabits its own Umwelt – a subjective perception of the physical world shaped by its sensory and motor capacities. For foveate primates such as the common marmoset, vision captures only the frontal space, leaving the vast rear hemisphere to be monitored acoustically. How does auditory cortex meet this ecological challenge, and how is its code influenced by visual events that compete for
attention in the frontal field?
Methods and Results: We recorded single units in the auditory cortex of awake marmosets in response to 200-ms broadband noises presented from loudspeakers spanning –180° to +180° in azimuth. In alternate sessions a wide-field visual image (±40° azimuth) was presented at 0°, providing a frontal visual stimulation. Eye position was continuously tracked with an infrared eye-tracking system. Spatial tuning curves were analyzed with circular statistics to estimate best azimuth and tuning width.
Three principles emerged. (1) Direction–sharpness coupling. Tuning sharpened progressively toward the contralateral pole, culminating in a dense cluster of narrowly tuned neurons at -90°. Fisher-information analysis revealed that this non-uniform distribution maximizes directional precision at midline 0°– exactly where gaze can be rapidly deployed. (2) Vision-dependent gain without remapping. The frontal wide-field image multiplicatively scaled firing rates yet left best azimuth unchanged, revealing a cross-modal gain field rather than a spatial realignment. (3) Rear-space enhancement. Gain modulation of firing rates was strongest in the contra-rear field, suggesting that visual stimulation in front strengthens monitoring of the invisible rear hemifield.
Conclusions: Taken together, these results outline a dual strategy for constructing a 360° auditory Umwelt: stable directional “anchors” anchored by contralateral sharp-tuned neurons, and dynamic gain control that boosts regions left outside the visual fovea when frontal space becomes behaviorally loaded. Our findings bridge classic spatial-hearing theory with Umwelt thinking and offer a neural explanation for how primates maintain situational awareness while looking forward.
Multiscale temporal integration in the auditory cortex
Sam Norman-Haignere, University of Rochester Medical Center
Abstract: Natural sounds such as vocalizations and speech are organized across many different timescales, each of which is temporally variable and context dependent. The auditory system must therefore have mechanisms for flexibly integrating information across multiple timescales. Yet the principles that underlie how temporal information is integrated over time, and how adaptable these timescales are, remain poorly understood. To address this challenge, we have developed an experimental method to measure integration windows from nonlinear systems such as the brain. The temporal context invariance (TCI) method measures responses to sequences of sound segments, surrounded by different context segments, and estimates the smallest segment duration needed to produce a context-invariant response. We have applied the TCI method to responses from computational models, including speech-trained deep neural networks (DNNs), in order to validate the approach, understand how AI systems learn to integrate information in natural sounds, and develop testable predictions for neural experiments. We found that training DNNs to recognize information from speech causes the networks to learn hierarchically organized windows that increase across network layers and also become more flexible, dynamically expanding and contracting as speech is time compressed and stretched. We found that units in the auditory cortex of the ferret show clear evidence of a time-limited integration window, beyond which stimuli have little influence on the neural response, and that this window varies substantially across the neural population. Anatomically, integration windows showed clear evidence of hierarchical organization with substantially longer integration windows in non-primary regions. However, we did not find evidence that integration windows dynamically expanded or contracted as sounds are time-compressed or stretched. Similar results were observed in the human auditory cortex using spatiotemporally precise intracranial recordings from neurosurgical patients. If time allows, we will discuss current and future work investigating how listeners perceptually adapt to changes in speech rate over the timescale of minutes, and potential neural correlates of this adaptation in the brain, as well as whether neural integration windows vary with the behavioral state of animals.
SPATIAL NAVIGATION
Spatial Representations Emerge in a Model Linking Global Maps to First-Person Perspectives
Jeff Krichmar, UC Irvine
Abstract: Humans and many animals possess the remarkable ability to navigate environments by seamlessly switching between first-person perspectives (FPP) and global map perspectives (GMP). However, the neural mechanisms that underlie this transformation remain poorly understood. In this study, we developed a variational autoencoder (VAE) model, enhanced with recurrent neural networks (RNNs), to investigate the computational principles behind perspective transformations. Our results reveal that temporal sequence modeling is crucial for maintaining spatial continuity and improving transformation accuracy when switching between FPPs and GMPs. The model's latent variables spontaneously develop spatial representations similar to those seen in the distributed cognitive maps of the mammalian brain, such as head direction cells, place cells, corner cells, and border cells - but notably not grid cells - suggesting that perspective transformation engages multiple brain regions beyond the hippocampus and entorhinal cortex. Furthermore, our findings demonstrate that landmark encoding, particularly proximal environmental cues such as boundaries and objects, play a critical role in enabling successful perspective shifts, whereas distal cues are less influential. These insights on perspective linking provide a new computational framework for understanding spatial cognition and offer valuable directions for future animal and human studies. It highlights the contributions of temporal sequences, distributed representations, and proximal cues for navigating complex environments.
Structured remapping in the subiculum: A shared latent code underlies spatial transformations across environments
Dori Derdikman, Technion - Israel Institute of Technology
Abstract: "How does the brain preserve a coherent sense of space across distinct environments?
In this study, we investigated how neurons in the subiculum, a key output region of the hippocampal formation, encode spatial information as animals explore different rooms.
Using large-scale Neuropixels recordings from freely moving mice, we observed that while individual neurons frequently shifted their firing locations, the population-level activity exhibited smooth, structured transformations of the spatial map.
To probe the principles underlying this stability, we developed a data-driven model that decodes 2D position from subicular activity through a shared latent representation. Temporal dynamics are captured by a recurrent network (RNN), and environment-specific position estimates were produced through dedicated decoders. This structure allowed us to infer the transformation between spatial representations across environments by analyzing how the shared latent space was mapped onto position in each context. Despite its capacity to learn complex mappings, we observed that transformations between environments were often well-approximated by simple affine operations, such as rotations, scalings, and translations. These findings reveal that beneath the complexity of neural remapping lies a low-dimensional, structured transformation, suggesting that the brain maintains spatial coherence across contexts through flexible yet geometrically consistent coding strategies."
Path integration gain recalibration and theta rhythm
James Knierim, Johns Hopkins University
Abstract: Path integration (PI) is a computational process in which an organism’s velocity vector is integrated over time to update its location. The spatial firing of place cells and grid cells are driven in large part by PI, in conjunction with powerful influences of landmarks. Prominent models of place and grid cells posit a bump of activity on a sheet attractor, with the movement of the bump driven by the animal’s movement through the world. Accurate PI requires a fine-tuned gain factor that determines the magnitude of bump movement on the attractor sheet based on the magnitude of animal movement through the world. We previously used virtual reality to show that this “PI gain” is a plastic variable that the system learns through feedback from landmarks (Jayakumar et al., 2019). To understand network mechanisms behind the PI-gain calibration process, we now investigated how PI gain affects the dynamics of the hippocampal theta rhythm. We placed PI in conflict with landmarks by moving the distal landmark array by different amounts in a feedback loop based on the animal’s movement. Although theta phase precession (occurring in early theta phases) was unaffected in the landmark frame of reference, phase procession (occurring in later theta phases) was weakened substantially with increasing conflict between PI and landmark cues (Sueoka et al., 2025). Theta frequency in the LFP decreased with increasing conflict between PI and landmark cues, and the frequency began to return to baseline levels as the PI gain was recalibrated (Park et al., unpublished). We adapted a computational model by Chu et al. (2023) to show that these effects can be explained by decreases in the excitatory drive onto a ring attractor network caused by the mismatch between the PI and landmark cues. The decrease in phase procession supports models suggesting that the early phase of theta is involved in new learning (Hasselmo, 2005). Furthermore, the changes in theta LFP frequency provide a critical, physiological readout of the current value of the PI gain, an otherwise hidden variable that is obscured by the overriding influence of the external landmarks.
Path integration impairments reveal early cognitive changes in Subjective Cognitive Decline
Zoran Tiganj, Indiana University
Abstract: Path integration, the ability to track one’s position using self-motion cues, is critically dependent on the grid cell network in the entorhinal cortex, a region vulnerable to early Alzheimer’s disease pathology. In this study, we examined path integration performance in individuals with subjective cognitive decline (SCD), a group at increased risk for Alzheimer’s disease, and healthy controls using an immersive virtual reality task. We developed a Bayesian computational model to decompose path integration errors into distinct components. SCD participants exhibited significantly higher path integration error, primarily driven by increased memory leak, while other modelling-derived error sources, such as velocity gain, sensory and reporting noise, remained comparable across groups. Our findings suggest that path integration deficits, specifically memory leak, may serve as an early marker of neurodegeneration in SCD and highlight the potential of self-motion-based navigation tasks for detecting pre-symptomatic Alzheimer’s disease-related cognitive changes.
Recalibration of hippocampal path integration gain without angular head velocity gain recalibration of head direction cells
Yotaro Sueoka, Johns Hopkins University
Abstract: Path integration (PI) is a process in which animals track their location by integrating self-motion cues. The PI gain of CA1 place cells is a key parameter that converts distance traveled in the world to corresponding distance traveled in the cognitive map. This gain recalibrates to new values when PI is placed in conflict with visual landmark navigation. However, the locus of PI gain plasticity is unknown.
A key input to the place cells is the head direction (HD) system, which orients the cognitive map. To test whether HD cells also exhibit PI gain recalibration, we recorded from place cells and thalamic HD cells while 4 Long-Evans rats ran around a circular track inside a planetarium-style VR apparatus. An array of visual landmarks was moved in the same or opposite direction as the rat’s movement, causing an illusion that the rat was moving slower (or faster) than it actually was. The motion of the landmarks was controlled by an experimental gain, G, which specified the animal’s speed in the landmark frame relative to the lab frame. In 29/39 sessions, the tuning of both place cells and HD cells remained anchored to the moving landmarks. Once the landmarks were turned off, the rate of update of place-cell activity, now solely driven by self-motion cues, was strongly correlated with the preceding G, demonstrating PI gain recalibration. Simultaneously recorded HD cells exhibited identical gains as the hippocampal gain.
The synchronized recalibration could result from plasticity occurring either i) in the angular head velocity (AHV) inputs onto the HD system (feedforward model), or ii) downstream of the HD system, which entrains the HD activity (feedback model). To dissociate these hypotheses, we analyzed head scanning behaviors, characterized by lateral head movements with minimal changes in body position. Given the absence of locomotion, HD cell activity is driven primarily by the AHV inputs. Strikingly, the HD gain during head scans in the absence of landmarks was close to 1, with no correlation to G (p = 0.515). The lack of recalibration during head scans supports the feedback model of PI gain recalibration under these experimental conditions.
Explicit error coding can mediate gain recalibration in continuous bump attractor networks
Noah Cowan, Johns Hopkins University
Abstract: Continuous bump attractor networks (CBANs) are a prevailing model for how neural circuits represent continuous variables. CBANs maintain these representations by temporally integrating inputs that encode differential (i.e., incremental) changes to a given variable. The accuracy of this computation hinges on a precisely tuned integration gain. Experiments have shown that the brain can recalibrate this gain using ground-truth sensory information, yet existing CBAN models rely on biologically implausible or currently unknown plasticity rules for recalibration. Here, we demonstrate that ring-type CBANs can recalibrate their integration gain through two mechanisms that rely on well-established, biologically plausible forms of plasticity. In the first mechanism, the spatially distributed synapses conveying incremental information to the attractor are plastic, allowing the integration gain to become transiently inhomogeneous during recalibration. In the second, plasticity is implemented in other components of the network, keeping the gain homogeneous during recalibration. Both mechanisms require explicit error signals that drive plasticity. We instantiate each mechanism within a CBAN, demonstrating their potential for biologically plausible, adaptive coding of continuous variables.
Sparse-to-dense coding transformation between hippocampal subregions CA3 and CA1 in very large environments
Nachum Ulanovsky, Weizmann Institute of Science
Abstract: The mammalian hippocampus is crucial for spatial memory and navigation. It contains place-cells, spatially-selective neurons which reside in two distinct hippocampal subregions with dramatically-different anatomical connectivity: areas CA1 and CA3. Prior studies have found surprisingly similar spatial coding between CA1 and CA3 place-cells, consisting mainly of single place-fields with similar field sizes. This raises a conundrum: Why would two consecutive subregions with major structural differences exhibit identical neural coding? Importantly, all prior comparisons of CA1 and CA3 were conducted in animals navigating in relatively small arenas. Here, we hypothesized that different neural coding will be revealed in large naturalistic environments — and we tested this hypothesis by simultaneously recording from CA1 and CA3 neurons in bats flying in very large environments: flight-tunnels up to 200-meters long. We discovered dramatically distinct neuronal coding between CA1 and CA3 place-cells: While CA1 neurons exhibited dense spatial coding, consisting of multiple place-fields, CA3 neurons exhibited ultra sparse spatial coding, consisting predominantly of single place-fields. Despite the striking difference in the number of place-fields, the sizes of individual place-fields were remarkably similar between the two subregions, across 5 different environment sizes ranging between 6–200 meters. Using a neural-network model, we showed that the sparse-to-dense architecture can facilitate fast learning of new spatial maps in the hippocampus. Additionally, we found that in a large multi-compartment environment, place-cells in both subregions exhibited strong contextual effects of trajectory history (‘retrospective coding’), which could last for >100 meters and >12 seconds. The retrospective coding was more robust in CA1 than in CA3. Together, by using large naturalistic environments we revealed a surprising CA3-to-CA1 coding transformation, which serves to reformat spatial information into a more efficient neural code for readout by the cortex.
VISION
Looking at Neural Representation of Translucent Object Appearance in the Macaque Brain through Artificial Neural Network
Ryusuke Hayashi, National Institute of Advanced Industrial Science and Technology (AIST)
Abstract: The degree to which light passes through a material gives rise to the perception of translucency. This visual attribute serves as an important cue for evaluating properties such as the juiciness of food or the condition of skin, making it essential in daily life. However, the neural substrates underlying the perception of translucent appearance remain largely unknown. The study of translucency perception is particularly challenging because it depends on a complex interplay of optical parameters and configural factors, many of which are still poorly understood.
In this study, we generated object images varying in shapes and degrees of translucency using the data-driven Translucent Appearance Generation (TAG) model (Liao et al., 2023), which encodes object shape and translucency in latent variables through unsupervised generative learning from a wide variety of real-world translucent object images. Using this image set, we conducted human behavioral experiments, macaque electrophysiological recordings, and in-silico experiments.
We implanted four micro-electrode arrays (512 electrodes in total) on the cortical surface along the macaque ventral visual pathway, specifically the inferior temporal (IT) cortex, and recorded population neural activity during object image presentation. We then compared this neural population activity with translucency ratings obtained from human behavioral experiments. Representational similarity analysis revealed a significant correlation between the human evaluation data and the macaque IT responses, whereas no such correlation was observed for pixel-scrambled versions of the images. Furthermore, Linear Discriminant Analysis trained on neural data from translucent images generalized well to novel test images of translucent objects, with classification outcomes aligning with human translucency ratings. In contrast, a popular artificial neural network designed to evaluate image quality in line with human perception failed to show such alignment. These findings demonstrate that neural population activity in the macaque IT cortex encodes the translucent appearance of objects in a manner that cannot be explained as a straightforward outcome of a simple object recognition neural network.
Modeling dorsal and ventral visual pathways for active vision
Zhongming Liu, University of Michigan
Abstract: The human visual system employs two parallel streams: a dorsal pathway for spatial analysis and action guidance, and a ventral pathway for object recognition. Inspired by this architecture, we have developed dual-stream neural network models that integrate retinal sampling, attention-driven eye movements, and recurrent processing. These models demonstrate that retinal transformation and sequential fixations naturally give rise to human-like gaze patterns and enhance robustness against adversarial perturbations. Comparisons with fMRI during natural movie viewing further show that dual-stream networks reproduce the functional specialization of dorsal and ventral visual pathways in the human brain. Extending this framework, we show that functional specialization can also emerge from structurally segregated pathways trained with self-supervised learning on unlabeled videos. One stream learns to represent objects, while the other learns potential actions applied to them. Together, they enable not only recognition of ongoing input but also anticipation of future states. This line of work illustrates how predictive learning, retinal transformations, and dual-stream segregation converge to explain robust and specialized cortical computation, while also advancing the design of adaptive and resilient artificial vision systems.
Robustness and Flexibility of Visual System Organization and Function
Herwig Baier, Max Planck Institute for Biological Intelligence
Abstract: The visual system of zebrafish has become one of the best studied preparations in behavioral and systems neuroscience. Around 40 molecularly and morphologically defined types of retinal ganglion cells (RGCs) serve as matched filters for behaviorally relevant stimulus features and feature configurations, including background lighting, optic flow, prey items, and objects on a collision course. RGCs transmit their signals via axon projections to a well-defined set of areas in the hypothalamus, prethalamus, thalamus, pretectum, and tectum. The major visuomotor hub, the optic tectum, harbors nine RGC input layers. Each of these layers recombines information of multiple single features, such as object size, expansion dynamics, and direction of motion. The retinotopic map in the tectum is locally adapted to visual scene statistics and visual subfield-specific behavioral demands. Tectofugal projections to premotor centers in hindbrain and tegmentum are topographically organized by the nature and strength of the behavioral commands that they convey. Each visually evoked response investigated so far, such as prey capture, loom avoidance, or shoaling, is controlled by a dedicated multi-station pathway, which –at least in the larva– resembles a labeled line.
Although this neural architecture is to a large extent hardwired and develops in the absence of visual experience, there is substantial flexibility in how visual stimuli are processed in the larval fish brain:
• Internal states, such as hunger, stress, or fear, modulate behavioral responses via monoaminergic and neuropeptidergic projections to the tectum.
• Specific neural circuits, particularly an isthmotectal loop, allow stimulus prioritization and thus serve an elementary form of spatial attention.
• Efference copies allow the subtraction of self-motion, such that the fish can process two identical visual stimuli in distinct ways.
• Prior experience facilitates visual responses to prey and improves hunting success.
In conclusion, natural selection has shaped a robust, yet flexible, visual system that enables adaptive behavior. The organizational principle is that of multiple independent-yet-interacting parallel neural pathways. The rapidly growing knowledge of neuronal cell types and synapse-scale connectivity in the zebrafish brain provides an opportunity to discover the cellular and synaptic basis of elementary perceptual and cognitive functions in the vertebrate nervous system.
Cortical processing of high-acuity vision
Daniel Butts, University of Maryland, College Park
Abstract: "When we want to see something, we look at it. Our highest acuity vision is present in the central one degree of our visual field (the fovea), and eye movements direct the fovea to regions of interest in the visual scene. Our visual perception relies heavily on foveal processing and its interactions with eye movements: only by artificially holding our gaze fixed can we notice that most of our vision is blurry.
Despite its importance for human vision, cortical neurons representing the fovea are essentially unstudied, because tiny eye movements — smaller than the resolution of traditional eye tracking — make it impossible to know the visual stimulus with sufficient spatial precision. As a result, nearly all studies of visual selectivity of neurons in the primary visual cortex (V1) are based on recordings outside the fovea, with the assumption that foveal V1 neurons perform the same processing of color and form as parafoveal and peripheral V1 neurons, but at smaller spatial scales.
Here, we combine large-scale neurophysiology with computational modeling to perform the first detailed characterizations of spatial and color processing of foveal V1 neurons. We infer the precise eye position at each time point by leveraging the exquisite spatial sensitivity of foveal neurons, using their population activity to determine the most likely position of the stimulus on the retina at each moment. Such neurophysiological eye tracking yields accuracy down to 1 arc-minute (1/60 of a degree), and allows us to finally determine the degree to which foveal processing of color and form resembles what we know about the rest of V1 from half a century of visual neurophysiology.
In fact, we have found that fovea representations are distinct, with an apparent division between the processing of color-independent form at the highest spatial resolution, and lower-resolution processing of color. This division, and many of the detailed spatio-chromatic properties of foveal V1 neurons, are consistent with the constraints implicit in sampling from the cone mosaic at the highest resolutions. These measurements not only fix the upper limits of spatial and color vision, but are also likely critical in shaping visual behavior."
SENSORI-MOTOR / DECISION MAKING
Data-Driven Feature Extraction and Stability of ECoG Speech and Hand Motor Decoding in an ALS Patient
Dean J. Krusienski, Virginia Commonwealth University
Abstract: Recent studies have shown significant promise toward the development of speech neuroprostheses using intracranial signals. These advances have made chronic neural recording in clinical populations both feasible and informative for long-term decoding studies. As an extension of our intracranial data collected under the CRCNS award, we examine approximately 2.5 years of neural data from a clinical trial involving an individual with progressive ALS, performing established speech and hand grasp tasks. The participant was implanted with two electrocorticographic (ECoG) arrays over the sensorimotor cortex, targeting regions associated with speech and upper-limb motor control. Such rare longitudinal ECoG data from an ALS patient provides the unique opportunity to evaluate signal and decoding stability over time, as well as train data-hungry deep-learning models for data-driven feature characterization. Preliminary analysis indicates that bandpower in the high beta (21-30 Hz) range provides more stable decoding performance over time compared to the conventional high gamma (70-170 Hz) for both hand grasp and speech activity detection. To further explore the relevant feature space, we employed the well-established EEGNet deep-learning architecture for neural signals. Individual EEGNet models were trained to classify action versus idle states using raw ECoG data from grasp and speech tasks, both independently and jointly. The learned convolutional filters were examined to identify interpretable spectral and spatial features associated with each modality. The resulting filters revealed some expected spatio-temporal patterns including activity resembling local motor potentials over hand areas for the hand movements and broadband gamma over more diffuse regions for speech production. Additionally, unexpected patterns were observed that warrant further investigation, such as low-frequency ventral motor activations for hand movements and distinct gamma-range spectral peaks for the combined models. These findings aim to elucidate the shared and distinct neural substrates underlying speech and grasp tasks, while providing new insights into the spatial and spectral characteristics of relevant features for neural decoding.
CLAWing toward threshold: how the dynamic interplay of cortico-basal ganglia-thalamic pathways shapes the decision-making process
Timothy Verstynen, Carnegie Mellon University
Abstract: Significant effort has been spent looking at specific roles that specific neural populations in the cortico-basal ganglia pathways play in guiding decisions and learning from them. In this study, we take on the much less investigated questions of how activity evolves within the whole circuit during the course of individual decisions, the nature of variability within this process, and how this dynamic evolution changes with learning. To address these questions, we developed and used a novel computational framework that we call CLAW (Circuit Logic Assessed via Walks) to trace the instantaneous flow of neural activity as it progresses through stochastic, spiking cortico-basal ganglia-thalamic (CBGT) networks engaged in a virtual decision-making task. Our results uncover distinct deliberation and commitment phases characterized by which CBGT components are dominant and suggest how differences in subpopulation dominance underlie distinct decision strategies. Moreover, we show how dopamine-dependent synaptic plasticity based on decision outcomes can alter the nature of CBGT activity during decisions and shift the likelihood of decision strategies from deliberation to reward-directed commitment. Finally, we map these results to the drift-diffusion model (DDM) framework, recasting them in terms of gradual changes in DDM parameters that provide an algorithmic interpretation of how learning leads to improved decision speed and accuracy while preserving the capacity for caution.
The Weight of an Error: Generalized Information Integration through Weighted Prediction Errors
Alireza Soltani, Dartmouth College
Abstract: "Many naturalistic learning tasks involve evaluating and integrating information sampled from stochastic environments. For instance, in foraging tasks, individuals estimate the probability of finding rewards after repeatedly visiting different locations; and bandit tasks involve estimating the average payouts after repeatedly gambling at different slot machines. Although the cognitive processes underlying evidence integration have been widely studied, less attention has been devoted to studying how and why individuals integrate information in different ways. Understanding the neural and cognitive bases of individual variability in evidence integration can significantly enhance our overall understanding of learning and decision making overall.
In this study, we examined the mechanisms of evidence integration using mathematical and neural network models applied to three different sets of behavioral data. We propose a framework for how agents flexibly incorporate new information into their existing estimates using prediction error weighted by the amount of information previously acquired. We show that this simple model can reproduce a variety of strategies for evidence integration, then implement this model in a biologically plausible neural network that closely resembles well-studied reward-learning systems in the brain. First, we demonstrate individual differences in evidence integration by analyzing human behavioral data from three sequential estimation tasks. We then fit our model to each individual and identify the cognitive parameters that account for the observed heterogeneities.
Our model captures these diverse strategies and can reproduce both the recency bias and optimal strategic integration. Critically, we show how variability within trials, between trials, and across individuals can be linked to biological parameters in our neural network model, including the number of neurons and the pattern of synaptic connectivity. Furthermore, we demonstrate that our simulated neurons replicate activity patterns in rostrolateral PFC and dorsal ACC related to prediction error and social weighting.
These results suggest that our framework and its neural network implementation reveal plausible, flexible, and resource-efficient mechanisms for evidence integration. Specifically, humans may rely on weighted error-driven learning, rather than more resource-intensive operations such as Bayesian inference, when making estimates about sequences of data.
Dynamic, distributed decision-making by frontostriatal circuits making
Rishidev Chaudhuri, University of California, Davis
Abstract: The ability to integrate sensory evidence over time and then use it to drive a later decision is key to flexible behavior. This flexible decision-making relies on a distributed network of multiple cortical and subcortical areas working in concert. We combined large-scale multi-region recordings with latent variable and network models to characterize the dynamics and inter-areal interactions of frontal and striatal circuits while rats perform an integration of evidence task. We found that neural activity follows stereotyped low-dimensional but nonlinear paths in state-space over the course of a trial, suggesting dynamic coding by internally-generated neural trajectories rather than static or predominantly stimulus-driven representations. Cortex and striatum are tightly coupled, but cortex more strongly reflects changes in task epoch, state, and performance, suggesting that in our tasks the basic anatomical unit of computation is a flexibly modulated cortico-striatal loop. Finally, behavior is not homogenous and instead is well-explained by rats switching between a small number of latent states. These latent states differ in their degree of task engagement and affect both behavior and underlying network dynamics. Our results thus point to a highly-dynamic and distributed decision-making circuit that is strongly shaped by internal state.
Perturbing the brain from the inside out
Paul Nuyujukian, Stanford University
Abstract: The majority of motor systems studies, particularly in large animals, have been correlational in na- ture, motivating the need for novel studies to explore the generalization and causality of neuronal low-dimensional states. In this presentation, I will share new findings that challenge our understand- ing of dimensionality of motor cortex, with both lower ( 2) and minimum upper bounds ( 200), as uncovered by order 102 and 103 simultaneous electrode recordings from chronically implanted, large animals across conventional and freely-moving settings. To define the lower bound, we show that long-term (months and years) decoding stability of a very low-dimensional (i.e., 2-3) subspace un- covered via multiclass linear discriminant analysis matches or exceeds the performance (Abdulla, et al., under review) of a decoder regressed against a CCA-aligned 10-dimensional subspace identified by PCA (Gallego, et al., Nature Neuro 2020). The minimum upper bound was found by tracking the dimensions needed to explain 80% of the variance of PCA-reduced spiking neuronal thresholds from 1024 chronically implanted electrodes in unconstrained large animals performing both a struc- tured 2D-reaching task. Moreover, we found that the relationship between the number of recorded electrodes and estimated dimensionality followed a power law with an exponent of 0.9 (Silvernagel, et al., under review). Similarly, if the PC subspace for a given multi-hour reaching session is re- calculated every 16 trials, we find a rotating, but consistent and small number ( 5) of eigenvectors that are quickly and flexibly switched among, hinting that the brain may distribute motor control over a small, conserved, but rapidly switched set of redundant neural circuits. Further evidence supporting this finding is shown by the perturbation of the switching rate, and in some cases even the transient loss of, of these conserved eigenvectors after electrolytic lesioning, which is restored within approximately a week (Clarke, et al, under review). Taken together, these findings motivate careful reconsideration of the notions of motor neural latent dimensionality and stability in light of novel findings enabled by high-channel count recordings, unconstrained behavioral studies, and causal perturbations of the brain.
Dynamics of thalamocortical networks during sensory discrimination
Ariel Gilad, Hebrew University of Jerusalem
Abstract: "The cortex and thalamus are key players in sensory integration. While traditionally the thalamus was viewed as a passive relay of sensory input to the cortex, accumulating evidence suggests it actively contributes to higher order processing through dynamic, bidirectional interactions with cortical areas. A major challenge in characterizing thalamocortical loops is the heterogeneity within both structures, each comprising multiple subdivisions with potentially distinct functions and connectivity. Moreover, most studies focus on single sensory modalities, despite the likelihood that different modalities engage distinct thalamocortical subnetworks.
To address these gaps, we combined wide-field calcium imaging of the dorsal cortex with multi-fiber photometry targeting 13 thalamic nuclei, enabling simultaneous recording across cortex and thalamus in mice performing both auditory and tactile discrimination tasks. We found that each sensory modality engages distinct thalamocortical subnetworks that selectively encode task-relevant features. In parallel, several thalamic nuclei exhibited modality-invariant choice-related activity, suggesting their involvement in internal decision-related representations.
Directional network analysis revealed specific thalamocortical and corticothalamic interactions that were consistent across tasks, highlighting structured bidirectional information flow. Notably, these interactions differed across subnetworks, emphasizing the functional specificity within the broader thalamocortical system.
Together, our findings provide new insights into how distributed thalamocortical circuits contribute to multisensory decision-making and underscore the importance of simultaneously probing both structures across multiple tasks and modalities.
Saccade-Related Evoked Potentials and Their Role in Human Visual Encoding
Gansheng Tan, Washington University
Abstract: Saccade-related evoked potentials (EPs) reflect neural dynamics time-locked to eye movements and have been documented during natural vision in humans and other primates. Our previous studies found that saccade-related EPs could have two polarities in humans, evidenced by a bimodal distribution of the intertrial correlation between local field potential following saccades. We carefully validated these saccade-related EPs were not related to oculomuscular signals. In addition, the polarity of saccade-related EPs was associated with the phase of ongoing oscillations. These findings establish saccade-related EPs as a format of saccadic modulation. Building on this we asked whether saccade-related EPs contribute to visual encoding. Addressing this question is important because it provides insight into a mechanism for optimizing visual encoding during natural vision.
To test this, we recorded local field potentials and eye gaze data in invasively monitored epilepsy patients when performing a visual encoding task. We identified saccades and extracted saccade-related EP characteristics, including latency, amplitude, peak time, and the power and phase of preceding oscillations. We then used random forest classifiers to map the relationship between these EP characteristics and visual encoding performance tested the next day. We found that at channel-level, the balanced accuracy of the random forest model on held-out test data was 55.7%, exceeding all 200 values from a null distribution generated by label permutation. At the saccade level, we aggregated information across channels by averaging predicted probabilities, yielding a test ROC AUC of 0.662 and balanced accuracy of 62.6%, exceeding all values from the null distributions. These findings demonstrate that saccade-related EPs contribute to visual encoding, linking oculomotor-related neural dynamics to memory formation.
Interleaved Replay of Novel and Familiar Memory Traces During Slow-Wave Sleep Prevents Catastrophic Forgetting
Maksim Bazhenov, UC San Diego
Abstract: Humans and animals can learn continuously, acquiring new knowledge and integrating it into a pool of lifelong memories. Memory replay during sleep has been proposed as a powerful mechanism contributing to interference-free new learning. In contrast, artificial neural networks suffer from a problem called catastrophic forgetting, where new training damages existing memories. This issue can be mitigated by interleaving training on new tasks with past data; however, whether the brain employs this strategy remains unknown. In this work, we show that slow-wave sleep (SWS) employs an interleaved replay of familiar cortical and novel hippocampal memory traces within individual Up states of the slow oscillation (SO), allowing new memories to be embedded into the existing pool of cortical memories without interference. Using a combination of biophysical modeling and analyses of single-unit activity from the mouse retrosplenial cortex - for a mouse trained first in a highly familiar environment and then in a novel one - we found that hippocampal ripples arriving near the Down-to-Up or Up-to-Down transitions of the sleep SO can entrain novel memory replay, while the middle phase of the Up state tends to replay familiar cortical memories. This strategy ensures the consolidation of novel cortical memory traces into long-term storage while minimizing damage to familiar ones. Our study introduces idea of Structured Cortical Replay (SCoRe) - a framework explaining how memory traces acquired at different times during an animal’s lifespan are dynamically organized within the thalamo-cortico-hippocampal system to enable continual learning and offering insights into potential strategies for mitigating catastrophic forgetting in artificial networks.
MODELING / THEORY
Efficient Training of Large Networks with Constrained Feedback
Jonathan Kadmon, Hebrew University
Abstract: "Training modern artificial networks relies on backpropagating high-dimensional error signals—an algorithmic workhorse with little biological plausibility. Yet natural behaviors, from motor control to perceptual decisions, are governed by a few latent task variables and learned through sparse, indirect feedback, suggesting that precise, high-dimensional credit assignment may be unnecessary.
We present a theoretical and algorithmic framework showing that low-dimensional error feedback can match the accuracy of backpropagation while sharply reducing computational cost. Extending Feedback Alignment, we develop a theory of low-rank feedback that decouples forward and backward passes, constraining error signals through a low-dimensional bottleneck while preserving rich feedforward representations. Formal theoretical analysis of linear networks proves that large systems learn efficiently from low-dimensional errors if—and only if—feedback weights adapt to the error subspace, yielding local learning rules that operate with narrow feedback pathways yet achieve backprop-level performance.
Guided by the theory, we design scalable algorithms that cut training cost without loss of accuracy across diverse architectures, including deep convolutional and transformer networks. For visual transformers, training FLOPs drop by 30–40 %, with larger models achieving even greater savings.
Restricting feedback dimensionality also imprints distinctive geometric signatures on learned representations, offering testable predictions for neural circuits with anatomically constrained feedback—a direct bridge between AI algorithms and brain function.
Our results show that effective credit assignment need not scale with network size, challenging conventional practice and opening two avenues: (i) a biologically grounded route to more efficient large-scale AI and (ii) a principled framework for probing how low-dimensional feedback shapes cortical representations.
Causality as the Minimum Energy Principle
Moo Chung, University of Wisconsin-Madison
Abstract: Causal inference is essential for understanding how brain regions influence one another to support cognition and behavior. In the brain, information is rarely transmitted in a simple feedforward manner; instead, it emerges from the interplay of recurrent excitation, feedback inhibition, and long-range cyclic interactions. Traditional node-to-node models such as Granger causality, structural equation modeling (SEM) and dynamic casual modeling (DCM) are fundamentally limited to acyclic architectures and cannot capture cyclic or recurrent motifs, which are likely central to brain network dynamics. This limitation is particularly important for resting-state fMRI, where feedback and recurrent communication appear to stabilize and coordinate distributed systems even in the absence of tasks.
In a cohort of 400 resting-state fMRI subjects, we found that nearly half of brain network dynamics cannot be explained by simple feedforward interactions. Using a new model-free causal framework grounded in the minimum energy principle, we conceptualize causality as the flow of energy from higher-energy (cause) to lower-energy (effect) regions, with the system evolving toward minimal energy configurations. 54% of network energy reflects acyclic, feedforward transfer, whereas 45% arises from cyclic interactions. These cyclic interactions were dominated by two biologically interpretable motifs: (i) homotopic coupling across primary sensory and motor cortices, where bilateral symmetry of inputs and outputs makes interhemispheric coordination essential; and (ii) midline cerebello–limbic loops linking the vermis, cerebellar lobules, and olfactory/limbic structures, consistent with recurrent circuits that regulate arousal, autonomic function, and affective processes. Together, these results provide the first compelling evidence that recurrent and feedback pathways are not peripheral but central to the intrinsic organization of rs-fMRI networks.
Building on these insights, we introduce a framework for Counterfactual Causal Analysis that formulates interventions as energy perturbations on network flows. By simulating hypothetical disruptions of nodes, edges, or recurrent cycles, this approach predicts how causal organization would reconfigure under interventions that are infeasible in practice. This represents a major shift from purely descriptive to interventional neuroscience, enabling mechanistic insight and hypothesis generation for both basic brain science and clinical translation.
EEG-DaSh electroencephalography data and tool sharing resource
Arnaud Delorme & Oren Shriki, UC San Diego
Abstract: "We present EEG-DaSh (EEG-DataSharing), a new US–Israel data and tool sharing resource designed to accelerate discovery in human neuroelectromagnetic research. EEG-DaSh builds on the success of EEGLAB, OpenNeuro/NEMAR, and the Neuroscience Gateway to provide the first online platform offering curated, machine learning (ML) and deep learning (DL)–ready electroencephalography (EEG) and magnetoencephalography (MEG) datasets, together with open pipelines for analysis, modeling, and data augmentation. The archive will host data from over 25 laboratories worldwide (n ≈ 27,000 participants) and integrate datasets annotated using community standards such as BIDS and HED. Unlike existing repositories, EEG-DaSh provides preprocessed and labeled samples suitable for ML/DL, direct programmatic access via MATLAB/Python libraries, and free high-performance computing resources through the Neuroscience Gateway.
The platform is already online and actively used by the community. Notably, it serves as the foundation for an ongoing NeurIPS competition with more than 380 registered participants, underscoring its relevance to both neuroscience and AI researchers. Data are available without embargo, and users are provided with customizable pipelines for preprocessing, feature extraction, connectivity analysis, and generative modeling.
By aggregating and standardizing large-scale EEG/MEG data, EEG-DaSh aims to enable robust single-trial decoding, improve biomarker discovery, and foster collaborations across neuroscience, machine learning, and theoretical modeling communities. The project includes annual data releases, training workshops in the US and Israel, and Kaggle challenges to engage students and researchers. EEG-DaSh represents a major step toward open, reproducible, and large-scale brain data science and is positioned to become a widely used community resource.
Decomposing spiking neural networks with Graphical Neural Activity Threads
Bradley Theilman, Sandia National Lab
Abstract: To understand the computational capacities of the brain and how we might develop brain inspired AI algorithms, we need powerful abstractions for neural computation. Ideally, these abstractions should be naturally adapted to the spiking and synaptic dynamics of real brains. We present an alternative approach to analyzing spiking neural networks that avoids many of the implicit assumptions in current approaches for spiking network analysis and offers a route to new computational abstractions. Current approaches for building computational abstractions for spiking dynamics begin by sorting spikes into time bins and constructing population activity vectors that trace the dynamics of neural activity in a high dimensional space over time. While fruitful, these approaches necessarily smear out intrinsic relations between spikes and may obscure computationally-relevant features of neural dynamics. Our approach begins by constructing a directed acyclic graph directly from the synaptic relations between individual spikes. These synaptic relations must support the computations in the spiking network. The analysis combines spiking activity and the connectome into a unified mathematical object, without time bins. We show how this directed graph naturally decomposes into weakly connected subgraphs we call Graphical Neural Activity Threads (GNATs). These GNATs are well-defined and provide a picture of information flow through a spiking network. Furthermore, GNATs are defined by the relative timings between spikes and are thus robust to spike timing variations. I will then describe an algorithm that can efficiently find isomorphic GNATs in large spiking neural datasets. By identifying isomorphic GNATs, we identify putatively isomorphic computations. I will show how GNATs arising in the dynamics of spiking network models are constructed out of other GNATs, analogous to sampling in music production. Thus, GNATs exhibit compositionality. Because of their naturalness, robustness, and compositionality, GNATs provide a powerful basis for computational abstraction in spiking neural networks.
SYSTEMS
How cortical circuits talk: Reorganization of high-dimensional cortical activity across cognitive states and goal-directed behavior
Arseny Finkelstein, Tel Aviv University
Abstract: Cognitive processes unfold over diverse spatial and temporal scales, engaging activity across multiple cortical areas. Recent studies have begun to characterize spatiotemporal dynamics across the neocortex, yet how cortex-wide activity at the level of individual neurons supports behavior remains unclear. We addressed this using mesoscale calcium imaging of up to ~30,000 neurons simultaneously (~1,000,000 total) across more than 10 cortical regions, while mice performed a tongue-reaching task or during spontaneous activity.
Both local (intra-areal) and global (inter-areal) dynamics were high-dimensional, with variance broadly distributed across dimensions. During task engagement, dimensionality was transiently quenched around reward time across all regions, suggesting temporally coordinated dynamics. Global dynamics during spontaneous activity were even higher-dimensional than during the task, a difference that could reflect either 1) local dimensionality quenching or 2) enhanced inter-areal coupling during task performance. Consistent with the latter, we found that local dimensionality remained similar across behavioral states, despite largely distinct activity fluctuations, with only a small subset of dimensions shared between spontaneous and task-driven dynamics. Moreover, we identified high-dimensional communication subspaces linking cortical areas, with inter-areal communication strengthened during task performance. This suggests that while the overall complexity of neural dynamics within individual cortical areas remains stable across cognitive states, coordination across regions increases during goal-directed behavior.
Finally, task-related movements engaged a high-dimensional activity subspace with broad anatomical distribution, which interacted with communication subspaces in a distance-dependent manner, suggesting spatial gating of movement-related signals. Together, these findings reveal that distributed cortical activity and inter-areal communication reorganize in structured, high-dimensional subspaces to support goal-directed behavior.
Learning Resting State Dynamics From 40K FMRI Sequences
Vikas Singh, University of Wisconsin Madison
Abstract: We describe a dynamical-systems based model for resting-state functional magnetic resonance imaging (rs-fMRI), trained on a dataset of roughly 40K rs-fMRI sequences covering a wide variety of public and available-by-permission datasets. While most existing proposals use transformer backbones, we utilize multi-resolution temporal modeling of the dynamics across parcellated brain regions. We show that our model, MnemoDyn, is compute efficient and generalizes very well across diverse populations and scanning protocols. When benchmarked against current state-of-the-art transformer-based approaches, MnemoDyn consistently delivers superior reconstruction quality. Overall, we find that with such large-scale pre-training on (non-proprietary) rs-fMRI datasets, we get a highly performant model for various downstream tasks. Our results also provide evidence of the efficacy of the model on small sample size studies which has implications for various studies where resting state fMRI is a commonly acquired imaging modality.
Correspondence of large-scale functional brain network decline across aging mice and humans
Gagan Wig, The University of Texas at Dallas
Abstract: Human adult aging is accompanied by alterations in the organization of large-scale brain networks, in both health and disease. For example, increasing age is accompanied by reductions in the modularity of functional brain networks. These changes in network topology are captured by alterations in system segregation--a measure quantifying the extent to which distinct brain systems are functionally differentiated. In humans, declines in system segregation are accompanied by worse cognitive ability and alterations in brain function, are stratified by socio-environmental factors, and are prognostic of Alzheimer’s Disease dementia. However, the mechanisms underlying aging-related network reorganization have yet to be established. Establishing a cross-species model of brain network organization and decline across the lifespan could bypass methodological limitations in human lifespan research and allow for identification of the mechanisms and factors underlying age-associated brain network alterations.
Using densely sampled resting-state fMRI data acquired cross-sectionally and longitudinally in awake mice over a broad range of adulthood (n=52; 3-20 months), we describe organizational features and age-related differences of the mouse’s functional connectome. Mouse resting-state functional connectivity recapitulates known functional circuits, demonstrating the organizational validity of these signals. Graph theoretic analysis applied to functional connectivity reveals that mice exhibit modular architectures of functional brain network organization and that increasing age is associated with decreasing system segregation, indicative of network dedifferentiation analogous to observations in humans. Notably, mouse resting-state brain networks are more segregated than those of humans (determined using data from the Human Connectome Project and its developmental- and aging-counterparts (n=1179; 18-90 years)), reflecting a diminished proportion of long-range functional relationships that integrate distributed systems. Mice also exhibit slower rates of age-related decline in brain network organization relative to humans, highlighting important species differences in functional brain network organization and trajectories of brain network aging. These findings establish a large-scale model of functional brain network aging in mice and provide a translational bridge across species and spatial scales of analysis.
Submillisecond interactions in large-scale spike train recordings
Kamran Diba, University of Michigan Medical School
Abstract: We observed exquisite synchrony with an intricate temporal structure featuring sub-millisecond periods of precise excitation and inhibition in the cross-correlograms of spike times from hippocampal and neocortical recordings. This effect was observed both locally and at distances up to 500 microns, including between different types of neurons. Based on analyses, modeling and the striking temporal similarity with action potential waveforms, we propose that extracellular conduction of action potentials combined with neuron-to-neuron ephaptic coupling can account for these effects. These findings reveal potential mechanisms for extremely rapid neuronal computations in the mammalian brain.