Uncertainty estimation with prediction-error circuits

Nature Communications, Journal Year: 2025, Volume and Issue: 16(1)

Published: March 28, 2025

Neural circuits continuously integrate noisy sensory stimuli with predictions that often do not perfectly match, requiring the brain to combine these conflicting feedforward and feedback inputs according their uncertainties. However, how tracks both stimulus prediction uncertainty remains unclear. Here, we show a hierarchical prediction-error network can estimate positive negative neurons. Consistent prior hypotheses, demonstrate neural rely more on when are environment is stable. By perturbing inhibitory interneurons within circuit, reveal role in estimation input weighting. Finally, link our model biased perception, showing contribute contraction bias.

Language: Английский

---

Synaptic signaling modeled by functional connectivity predicts metabolic demands of the human brain

NeuroImage, Journal Year: 2024, Volume and Issue: 295, P. 120658 - 120658

Published: May 28, 2024

The human brain is characterized by interacting large-scale functional networks fueled glucose metabolism. Since former studies could not sufficiently clarify how these connections shape metabolism, we aimed to provide a neurophysiologically-based approach.

Language: Английский

---

A layered microcircuit model of somatosensory cortex with three interneuron types and cell-type-specific short-term plasticity

Cerebral Cortex, Journal Year: 2024, Volume and Issue: 34(9)

Published: Sept. 1, 2024

Abstract Three major types of GABAergic interneurons, parvalbumin-, somatostatin-, and vasoactive intestinal peptide-expressing (PV, SOM, VIP) cells, play critical but distinct roles in the cortical microcircuitry. Their specific electrophysiology connectivity shape their inhibitory functions. To study network dynamics signal processing to these cell cerebral cortex, we developed a multi-layer model incorporating biologically realistic interneuron parameters from rodent somatosensory cortex. The is fitted vivo data on cell-type-specific population firing rates. With protocol stimulation, responses when activating different neuron are examined. reproduces experimentally observed effects PV SOM cells disinhibitory effect VIP excitatory cells. We further create version short-term synaptic plasticity (STP). While ongoing activity with without STP similar, modulates Exc, presumably by changing dominant pathways. slight adjustments, also sensory recorded vivo. Our provides predictions involving can serve explore computational interneurons

Language: Английский

---

Decoding violated sensory expectations from the auditory cortex of anaesthetised mice: Hierarchical recurrent neural network depicts separate ‘danger’ and ‘safety’ units

European Journal of Neuroscience, Journal Year: 2022, Volume and Issue: 56(3), P. 4154 - 4175

Published: June 13, 2022

The ability to respond appropriately sensory information received from the external environment is among most fundamental capabilities of central nervous systems. In auditory domain, processes underlying this behaviour are studied by measuring auditory-evoked electrophysiology during sequences sounds with predetermined regularities. Identifying neural correlates ensuing novelty responses supported research in experimental animals. present study, we reanalysed epidural field potential recordings cortex anaesthetised mice frequency and intensity oddball stimulation. Multivariate pattern analysis (MVPA) hierarchical recurrent network (RNN) modelling were adopted explore these data greater resolution than previously considered using conventional methods. Time-wise generalised temporal decoding MVPA approaches revealed underestimated asymmetry between sound-level transitions paradigm, contrast tone changes. After training, cross-validated RNN model architecture four hidden layers produced output waveforms response simulated inputs that strongly correlated grand-average (r2 > .9). Units classified based on their properties characterised principal component sample entropy. These demonstrated spontaneous alpha rhythms, sound onset offset putative 'safety' 'danger' units activated relatively inconspicuous salient changes inputs, respectively. hypothesised existence corresponding biological sources naturally derived model. If proven, could have significant implications for prevailing theories processing.

Language: Английский

---

Recurrent Neural Network Model of Human Event-related Potentials in Response to Intensity Oddball Stimulation

Neuroscience, Journal Year: 2022, Volume and Issue: 504, P. 63 - 74

Published: Oct. 10, 2022

Language: Английский

---

Uncertainty-modulated prediction errors in cortical microcircuits

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: May 12, 2023

Abstract Understanding the variability of environment is essential to function in everyday life. The brain must hence take uncertainty into account when updating its internal model world. basis for are prediction errors that arise from a difference between current and new sensory experiences. Although error neurons have been identified layer 2/3 diverse areas, how modulates these learning is, however, unclear. Here, we use normative approach derive should modulate postulate represent uncertainty-modulated (UPE). We further hypothesise circuit calculates UPE through subtractive divisive inhibition by different inhibitory cell types. By implementing calculation UPEs microcircuit model, show types can compute means variances stimulus distribution. With local activity-dependent plasticity rules, computations be learned context-dependently, allow upcoming stimuli their Finally, mechanism enables an organism optimise strategy via adaptive rates.

Language: Английский

---

Localized estimation of event-related neural source activity from simultaneous MEG-EEG with a recurrent neural network

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2024, Volume and Issue: unknown

Published: March 4, 2024

Abstract Estimating intracranial current sources underlying the electromagnetic signals observed from extracranial sensors is a perennial challenge in non-invasive neuroimaging. Established solutions to this inverse problem treat time samples independently without considering temporal dynamics of event-related brain processes. This paper describes source estimation simultaneously recorded magneto- and electro-encephalography (MEEG) using recurrent neural network (RNN) that learns sequential relationships data. The RNN was trained two phases: (1) pre-training (2) transfer learning with L1 regularization applied layer. Performance scaled labels derived MEEG, magnetoencephalography (MEG), or electroencephalography (EEG) were compared, as results volumetric space free dipole orientation surface fixed orientation. Exact low-resolution tomography (eLORETA) mixed-norm L1/L2 (MxNE) methods also these data for comparison method. approach outperformed other terms output signal-to-noise ratio, correlation mean-squared error metrics evaluated against ground-truth field (ERF) potential (ERP) waveforms. Using MEEG fixed-orientation produced most consistent estimates. To estimate ERF ERP waveforms, generates within its internal computational units, driven by structure used training labels. It thus provides data-driven model transformations psychophysiological events into corresponding signals, which unique among MEG EEG reconstruction solutions.

Language: Английский

---

Modeling circuit mechanisms of opposing cortical responses to visual flow perturbations

PLoS Computational Biology, Journal Year: 2024, Volume and Issue: 20(3), P. e1011921 - e1011921

Published: March 7, 2024

In an ever-changing visual world, animals’ survival depends on their ability to perceive and respond rapidly changing motion cues. The primary cortex (V1) is at the forefront of this sensory processing, orchestrating neural responses perturbations in flow. However, underlying mechanisms that lead distinct cortical such remain enigmatic. study, our objective was uncover dynamics govern V1 neurons’ flow using a biologically realistic computational model. By subjecting model sudden changes input, we observed opposing excitatory layer 2/3 (L2/3) neurons, namely, depolarizing hyperpolarizing responses. We found segregation primarily driven by competition between external input recurrent inhibition, particularly within L2/3 L4. This division not L5/6 suggesting more prominent role for inhibitory processing upper layers. Our findings share similarities with recent experimental studies focusing influence top-down bottom-up inputs mouse during perturbations.

Language: Английский

---

Probing inter-areal computations with a cellular resolution two-photon holographic mesoscope

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: March 3, 2023

Brain computation depends on intricately connected yet highly distributed neural networks. Due to the absence of requisite technologies, causally testing fundamental hypotheses nature inter-areal processing have remained largely out-of-each. Here we developed first two photon holographic mesoscope, a system capable simultaneously reading and writing activity patterns with single cell resolution across large regions brain. We demonstrate precise photo-activation spatial temporal sequences neurons in one brain area while out downstream effect several other regions. Investigators can use this new platform understand feed-forward feed-back circuits precision for time.

Language: Английский

---

Predictive learning rules generate a cortical-like replay of probabilistic sensory experiences

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2023, Volume and Issue: unknown

Published: Feb. 18, 2023

Abstract The brain is thought to construct an optimal internal model representing the probabilistic structure of environment accurately. Evidence suggests that spontaneous activity gives such a by cycling through patterns evoked previous sensory experiences with experienced probabilities. brain’s emerges from internally-driven neural population dynamics. However, how cortical networks encode models into poorly understood. Recent computational and experimental studies suggest neuron can implement complex computations, including predictive responses, soma-dendrite interactions. Here, we show recurrent network spiking neurons subject same learning principle provides novel mechanism learn replay experiences. In this network, rules minimize probability mismatches between stimulus-evoked internally driven activities in all excitatory inhibitory neurons. This paradigm generates stimulus-specific cell assemblies remember their activation probabilities using within-assembly connections. Our contrasts statistical Markovian transition among assemblies. We demonstrate our well replicates behavioral biases monkeys performing perceptual decision making. results interactions intracellular processes dynamics are more crucial for cognitive behaviors than previously thought.

Language: Английский

---