Learning prediction error neurons in a canonical interneuron circuit

Desegregation of neuronal predictive processing

Neural circuits construct internal 'world-models' to guide behavior. The predictive processing framework posits that neural activity signaling sensory predictions and concurrently computing prediction-errors is a signature of those internal models. Here, to understand how the brain generates predictions for complex sensorimotor signals, we investigate the emergence of high-dimensional, multi-modal predictive representations in recurrent networks. We find that robust predictive processing arises in a network with loose excitatory/inhibitory balance. Contrary to previous proposals of functionally specialized cell-types, the network exhibits desegregation of stimulus and prediction-error representations. We confirmed these model predictions by experimentally probing predictive-coding circuits using a rich stimulus-set to violate learned expectations. When constrained by data, our model further reveals and makes concrete testable experimental predictions for the distinct functional roles of excitatory and inhibitory neurons, and of neurons in different layers along a laminar hierarchy, in computing multi-modal predictions. These results together imply that in natural conditions, neural representations of internal models are highly distributed, yet structured to allow flexible readout of behaviorally-relevant information. The generality of our model advances the understanding of computation of internal models across species, by incorporating different types of predictive computations into a unified framework.

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

PURPOSE

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

METHODS

51 healthy volunteers underwent simultaneous PET/MRI to obtain BOLD functional connectivity and [18F]FDG glucose metabolism. These multimodal imaging proxies of fMRI and PET were combined in a whole-brain extension of metabolic connectivity mapping. Specifically, functional connectivity of all brain regions were used as input to explain glucose metabolism of a given target region. This enabled the modeling of postsynaptic energy demands by incoming signals from distinct brain regions.

RESULTS

Functional connectivity input explained a substantial part of metabolic demands but with pronounced regional variations (34 - 76%). During cognitive task performance this multimodal association revealed a shift to higher network integration compared to resting state. In healthy aging, a dedifferentiation (decreased segregated/modular structure of the brain) of brain networks during rest was observed. Furthermore, by including data from mRNA maps, [11C]UCB-J synaptic density and aerobic glycolysis (oxygen-to-glucose index from PET data), we show that whole-brain functional input reflects non-oxidative, on-demand metabolism of synaptic signaling. The metabolically-derived directionality of functional inputs further marked them as top-down predictions. In addition, the approach uncovered formerly hidden networks with superior efficiency through metabolically informed network partitioning.

CONCLUSIONS

Applying multimodal imaging, we decipher a crucial part of the metabolic and neurophysiological basis of functional connections in the brain as interregional on-demand synaptic signaling fueled by anaerobic metabolism. The observed task- and age-related effects indicate promising future applications to characterize human brain function and clinical alterations.

Simple synaptic modulations implement diverse novelty computations

Detecting novelty is ethologically useful for an organism's survival. Recent experiments characterize how different types of novelty over timescales from seconds to weeks are reflected in the activity of excitatory and inhibitory neuron types. Here, we introduce a learning mechanism, familiarity-modulated synapses (FMSs), consisting of multiplicative modulations dependent on presynaptic or pre/postsynaptic neuron activity. With FMSs, network responses that encode novelty emerge under unsupervised continual learning and minimal connectivity constraints. Implementing FMSs within an experimentally constrained model of a visual cortical circuit, we demonstrate the generalizability of FMSs by simultaneously fitting absolute, contextual, and omission novelty effects. Our model also reproduces functional diversity within cell subpopulations, leading to experimentally testable predictions about connectivity and synaptic dynamics that can produce both population-level novelty responses and heterogeneous individual neuron signals. Altogether, our findings demonstrate how simple plasticity mechanisms within a cortical circuit structure can produce qualitatively distinct and complex novelty responses.

Modeling circuit mechanisms of opposing cortical responses to visual flow perturbations

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

Fast adaptation to rule switching using neuronal surprise

In humans and animals, surprise is a physiological reaction to an unexpected event, but how surprise can be linked to plausible models of neuronal activity is an open problem. We propose a self-supervised spiking neural network model where a surprise signal is extracted from an increase in neural activity after an imbalance of excitation and inhibition. The surprise signal modulates synaptic plasticity via a three-factor learning rule which increases plasticity at moments of surprise. The surprise signal remains small when transitions between sensory events follow a previously learned rule but increases immediately after rule switching. In a spiking network with several modules, previously learned rules are protected against overwriting, as long as the number of modules is larger than the total number of rules-making a step towards solving the stability-plasticity dilemma in neuroscience. Our model relates the subjective notion of surprise to specific predictions on the circuit level.

The locus coeruleus as a global model failure system

Predictive processing models posit that brains constantly attempt to predict their sensory inputs. Prediction errors signal when these predictions are incorrect and are thought to be instructive signals that drive corrective plasticity. Recent findings support the idea that the locus coeruleus (LC) - a brain-wide neuromodulatory system - signals several types of prediction error. I discuss how these findings support models proposing that the LC signals global model failures: instances where predictions about the world are strongly violated. Focusing on the cortex, I explore the utility of this signal in learning rate control, how the LC circuit may compute the signal, and how this view may aid our understanding of neurodivergence.

Predictive learning by a burst-dependent learning rule

Humans and other animals are able to quickly generalize latent dynamics of spatiotemporal sequences, often from a minimal number of previous experiences. Additionally, internal representations of external stimuli must remain stable, even in the presence of sensory noise, in order to be useful for informing behavior. In contrast, typical machine learning approaches require many thousands of samples, and generalize poorly to unexperienced examples, or fail completely to predict at long timescales. Here, we propose a novel neural network module which incorporates hierarchy and recurrent feedback terms, constituting a simplified model of neocortical microcircuits. This microcircuit predicts spatiotemporal trajectories at the input layer using a temporal error minimization algorithm. We show that this module is able to predict with higher accuracy into the future compared to traditional models. Investigating this model we find that successive predictive models learn representations which are increasingly removed from the raw sensory space, namely as successive temporal derivatives of the positional information. Next, we introduce a spiking neural network model which implements the rate-model through the use of a recently proposed biological learning rule utilizing dual-compartment neurons. We show that this network performs well on the same tasks as the mean-field models, by developing intrinsic dynamics that follow the dynamics of the external stimulus, while coordinating transmission of higher-order dynamics. Taken as a whole, these findings suggest that hierarchical temporal abstraction of sequences, rather than feed-forward reconstruction, may be responsible for the ability of neural systems to quickly adapt to novel situations.

Short-term neuronal and synaptic plasticity act in synergy for deviance detection in spiking networks

Sensory areas of cortex respond more strongly to infrequent stimuli when these violate previously established regularities, a phenomenon known as deviance detection (DD). Previous modeling work has mainly attempted to explain DD on the basis of synaptic plasticity. However, a large fraction of cortical neurons also exhibit firing rate adaptation, an underexplored potential mechanism. Here, we investigate DD in a spiking neuronal network model with two types of short-term plasticity, fast synaptic short-term depression (STD) and slower threshold adaptation (TA). We probe the model with an oddball stimulation paradigm and assess DD by evaluating the network responses. We find that TA is sufficient to elicit DD. It achieves this by habituating neurons near the stimulation site that respond earliest to the frequently presented standard stimulus (local fatigue), which diminishes the response and promotes the recovery (global fatigue) of the wider network. Further, we find a synergy effect between STD and TA, where they interact with each other to achieve greater DD than the sum of their individual effects. We show that this synergy is caused by the local fatigue added by STD, which inhibits the global response to the frequently presented stimulus, allowing greater recovery of TA-mediated global fatigue and making the network more responsive to the deviant stimulus. Finally, we show that the magnitude of DD strongly depends on the timescale of stimulation. We conclude that highly predictable information can be encoded in strong local fatigue, which allows greater global recovery and subsequent heightened sensitivity for DD.

Local minimization of prediction errors drives learning of invariant object representations in a generative network model of visual perception

The ventral visual processing hierarchy of the cortex needs to fulfill at least two key functions: perceived objects must be mapped to high-level representations invariantly of the precise viewing conditions, and a generative model must be learned that allows, for instance, to fill in occluded information guided by visual experience. Here, we show how a multilayered predictive coding network can learn to recognize objects from the bottom up and to generate specific representations via a top-down pathway through a single learning rule: the local minimization of prediction errors. Trained on sequences of continuously transformed objects, neurons in the highest network area become tuned to object identity invariant of precise position, comparable to inferotemporal neurons in macaques. Drawing on this, the dynamic properties of invariant object representations reproduce experimentally observed hierarchies of timescales from low to high levels of the ventral processing stream. The predicted faster decorrelation of error-neuron activity compared to representation neurons is of relevance for the experimental search for neural correlates of prediction errors. Lastly, the generative capacity of the network is confirmed by reconstructing specific object images, robust to partial occlusion of the inputs. By learning invariance from temporal continuity within a generative model, the approach generalizes the predictive coding framework to dynamic inputs in a more biologically plausible way than self-supervised networks with non-local error-backpropagation. This was achieved simply by shifting the training paradigm to dynamic inputs, with little change in architecture and learning rule from static input-reconstructing Hebbian predictive coding networks.

A mechanism for differential control of axonal and dendritic spiking underlying learning in a cerebellum-like circuit

In addition to the action potentials used for axonal signaling, many neurons generate dendritic "spikes" associated with synaptic plasticity. However, in order to control both plasticity and signaling, synaptic inputs must be able to differentially modulate the firing of these two spike types. Here, we investigate this issue in the electrosensory lobe (ELL) of weakly electric mormyrid fish, where separate control over axonal and dendritic spikes is essential for the transmission of learned predictive signals from inhibitory interneurons to the output stage of the circuit. Through a combination of experimental and modeling studies, we uncover a novel mechanism by which sensory input selectively modulates the rate of dendritic spiking by adjusting the amplitude of backpropagating axonal action potentials. Interestingly, this mechanism does not require spatially segregated synaptic inputs or dendritic compartmentalization but relies instead on an electrotonically distant spike initiation site in the axon-a common biophysical feature of neurons.

Prediction of the Soil Permeability Coefficient of Reservoirs Using a Deep Neural Network Based on a Dendrite Concept

Changes in the pore water pressure of soil are essential factors that affect the movement of structures during and after construction in terms of stability and safety. Soil permeability represents the quantity of water transferred using pore water pressure. However, these changes cannot be easily identified and require considerable time and money. This study predicted and evaluated the soil permeability coefficient using a multiple regression (MR) model, adaptive network-based fuzzy inference system (ANFIS), general deep neural network (DNN) model, and DNN using the dendrite concept (DNN−T, which was proposed in this study). The void ratio, unit weight, and particle size were obtained from 164 undisturbed samples collected from the embankments of reservoirs in South Korea as input variables for the aforementioned models. The data used in this study included seven input variables, and the ratios of the training data to the validation data were randomly extracted, such as 6:4, 7:3, and 8:2, and were used. The analysis results for each model showed a median correlation of r = 0.6 or less and a low model efficiency of Nash–Sutcliffe efficiency (NSE) = 0.35 or less as a result of predicting MR and ANFIS. The DNN and DNN−T both have good performance, with a strong correlation of r = 0.75 or higher. Evidently, the DNN−T performance in terms of r, NSE, and root mean square error (RMSE) improved more than that of the DNN. However, the difference between the mean absolute percent error (MAPE) of DNN−T and the DNN was that the error of the DNN was small (11%). Regarding the ratio of the training data to the verification data, 7:3 and 8:2 showed better results compared to 6:4 for indicators, such as r, NSE, RMSE, and MAPE. We assumed that this phenomenon was caused by the DNN−T thinking layer. This study shows that DNN−T, which changes the structure of the DNN, is an alternative for estimating the soil permeability coefficient in the safety inspection of construction sites and is an excellent methodology that can save time and budget.

Interacting rhythms enhance sensitivity of target detection in a fronto-parietal computational model of visual attention

Even during sustained attention, enhanced processing of attended stimuli waxes and wanes rhythmically, with periods of enhanced and relatively diminished visual processing (and subsequent target detection) alternating at 4 or 8 Hz in a sustained visual attention task. These alternating attentional states occur alongside alternating dynamical states, in which lateral intraparietal cortex (LIP), the frontal eye field (FEF), and the mediodorsal pulvinar (mdPul) exhibit different activity and functional connectivity at α, β, and γ frequencies-rhythms associated with visual processing, working memory, and motor suppression. To assess whether and how these multiple interacting rhythms contribute to periodicity in attention, we propose a detailed computational model of FEF and LIP. When driven by θ-rhythmic inputs simulating experimentally-observed mdPul activity, this model reproduced the rhythmic dynamics and behavioral consequences of observed attentional states, revealing that the frequencies and mechanisms of the observed rhythms allow for peak sensitivity in visual target detection while maintaining functional flexibility.

Disinhibitory circuit mediated by connections from vasoactive intestinal polypeptide to somatostatin interneurons underlies the paradoxical decrease in spike synchrony with increased border ownership selective neuron firing rate

The activity of border ownership selective (BOS) neurons in intermediate-level visual areas indicates which side of a contour owns a border relative to its classical receptive field and provides a fundamental component of figure-ground segregation. A physiological study reported that selective attention facilitates the activity of BOS neurons with a consistent border ownership preference, defined as two neurons tuned to respond to the same visual object. However, spike synchrony between this pair is significantly suppressed by selective attention. These neurophysiological findings are derived from a biologically-plausible microcircuit model consisting of spiking neurons including two subtypes of inhibitory interneurons, somatostatin (SOM) and vasoactive intestinal polypeptide (VIP) interneurons, and excitatory BOS model neurons. In our proposed model, BOS neurons and SOM interneurons cooperate and interact with each other. VIP interneurons not only suppress SOM interneuron responses but also are activated by feedback signals mediating selective attention, which leads to disinhibition of BOS neurons when they are directing selective attention toward an object. Our results suggest that disinhibition arising from the synaptic connections from VIP to SOM interneurons plays a critical role in attentional modulation of neurons in intermediate-level visual areas.

Key role of neuronal diversity in structured reservoir computing

Chaotic time series have been captured by reservoir computing models composed of a recurrent neural network whose output weights are trained in a supervised manner. These models, however, are typically limited to randomly connected networks of homogeneous units. Here, we propose a new class of structured reservoir models that incorporates a diversity of cell types and their known connections. In a first version of the model, the reservoir was composed of mean-rate units separated into pyramidal, parvalbumin, and somatostatin cells. Stability analysis of this model revealed two distinct dynamical regimes, namely, (i) an inhibition-stabilized network (ISN) where strong recurrent excitation is balanced by strong inhibition and (ii) a non-ISN network with weak excitation. These results were extended to a leaky integrate-and-fire model that captured different cell types along with their network architecture. ISN and non-ISN reservoir networks were trained to relay and generate a chaotic Lorenz attractor. Despite their increased performance, ISN networks operate in a regime of activity near the limits of stability where external perturbations yield a rapid divergence in output. The proposed framework of structured reservoir computing opens avenues for exploring how neural microcircuits can balance performance and stability when representing time series through distinct dynamical regimes.

Rate and oscillatory switching dynamics of a multilayer visual microcircuit model

The neocortex is organized around layered microcircuits consisting of a variety of excitatory and inhibitory neuronal types which perform rate- and oscillation-based computations. Using modeling, we show that both superficial and deep layers of the primary mouse visual cortex implement two ultrasensitive and bistable switches built on mutual inhibitory connectivity motives between somatostatin, parvalbumin, and vasoactive intestinal polypeptide cells. The switches toggle pyramidal neurons between high and low firing rate states that are synchronized across layers through translaminar connectivity. Moreover, inhibited and disinhibited states are characterized by low- and high-frequency oscillations, respectively, with layer-specific differences in frequency and power which show asymmetric changes during state transitions. These findings are consistent with a number of experimental observations and embed firing rate together with oscillatory changes within a switch interpretation of the microcircuit.

On the relationship between predictive coding and backpropagation

Artificial neural networks are often interpreted as abstract models of biological neuronal networks, but they are typically trained using the biologically unrealistic backpropagation algorithm and its variants. Predictive coding has been proposed as a potentially more biologically realistic alternative to backpropagation for training neural networks. This manuscript reviews and extends recent work on the mathematical relationship between predictive coding and backpropagation for training feedforward artificial neural networks on supervised learning tasks. Implications of these results for the interpretation of predictive coding and deep neural networks as models of biological learning are discussed along with a repository of functions, Torch2PC, for performing predictive coding with PyTorch neural network models.

Prediction-error neurons in circuits with multiple neuron types: Formation, refinement, and functional implications

SignificanceAn influential idea in neuroscience is that neural circuits do not only passively process sensory information but rather actively compare them with predictions thereof. A core element of this comparison is prediction-error neurons, the activity of which only changes upon mismatches between actual and predicted sensory stimuli. While it has been shown that these prediction-error neurons come in different variants, it is largely unresolved how they are simultaneously formed and shaped by highly interconnected neural networks. By using a computational model, we study the circuit-level mechanisms that give rise to different variants of prediction-error neurons. Our results shed light on the formation, refinement, and robustness of prediction-error circuits, an important step toward a better understanding of predictive processing.

Visual association cortex links cues with conjunctions of reward and locomotor contexts

Postrhinal cortex (POR) and neighboring lateral visual association areas are necessary for identifying objects and interpreting them in specific contexts, but how POR neurons encode the same object across contexts remains unclear. Here, we imaged excitatory neurons in mouse POR across tens of days prior to and throughout initial cue-reward learning and reversal learning. We assessed responses to the same cue when it was rewarded or unrewarded, during both locomotor and stationary contexts. Surprisingly, a large class of POR neurons were minimally cue-driven prior to learning. After learning, distinct clusters within this class responded selectively to a given cue when presented in a specific conjunction of reward and locomotion contexts. In addition, another class contained clusters of neurons whose cue responses were more transient, insensitive to reward learning, and adapted over thousands of presentations. These two classes of POR neurons may support context-dependent interpretation and context-independent identification of sensory cues.

NMDA receptors in visual cortex are necessary for normal visuomotor integration and skill learning

The experience of coupling between motor output and visual feedback is necessary for the development of visuomotor skills and shapes visuomotor integration in visual cortex. Whether these experience-dependent changes of responses in V1 depend on modifications of the local circuit or are the consequence of circuit changes outside of V1 remains unclear. Here, we probed the role of N-methyl-d-aspartate (NMDA) receptor-dependent signaling, which is known to be involved in neuronal plasticity, in mouse primary visual cortex (V1) during visuomotor development. We used a local knockout of NMDA receptors and a photoactivatable inhibition of CaMKII in V1 during the first visual experience to probe for changes in neuronal activity in V1 as well as the influence on performance in a visuomotor task. We found that a knockout of NMDA receptors before, but not after, first visuomotor experience reduced responses to unpredictable stimuli, diminished the suppression of predictable feedback in V1, and impaired visuomotor skill learning later in life. Our results demonstrate that NMDA receptor-dependent signaling in V1 is critical during the first visuomotor experience for shaping visuomotor integration and enabling visuomotor skill learning.

Self-organization of a doubly asynchronous irregular network state for spikes and bursts

Cortical pyramidal cells (PCs) have a specialized dendritic mechanism for the generation of bursts, suggesting that these events play a special role in cortical information processing. In vivo, bursts occur at a low, but consistent rate. Theory suggests that this network state increases the amount of information they convey. However, because burst activity relies on a threshold mechanism, it is rather sensitive to dendritic input levels. In spiking network models, network states in which bursts occur rarely are therefore typically not robust, but require fine-tuning. Here, we show that this issue can be solved by a homeostatic inhibitory plasticity rule in dendrite-targeting interneurons that is consistent with experimental data. The suggested learning rule can be combined with other forms of inhibitory plasticity to self-organize a network state in which both spikes and bursts occur asynchronously and irregularly at low rate. Finally, we show that this network state creates the network conditions for a recently suggested multiplexed code and thereby indeed increases the amount of information encoded in bursts.

Deciphering how interneuron specific 3 cells control oriens lacunosum-moleculare cells to contribute to circuit function

The wide diversity of inhibitory cells across the brain makes them suitable to contribute to network dynamics in specialized fashions. However, the contributions of a particular inhibitory cell type in a behaving animal are challenging to untangle as one needs to both record cellular activities and identify the cell type being recorded. Thus, using computational modeling and theory to predict and hypothesize cell-specific contributions is desirable. Here, we examine potential contributions of interneuron-specific 3 (I-S3) cells - an inhibitory interneuron found in CA1 hippocampus that only targets other inhibitory interneurons - during simulated theta rhythms. We use previously developed multi-compartment models of oriens lacunosum-moleculare (OLM) cells, the main target of I-S3 cells, and explore how I-S3 cell inputs during in vitro and in vivo scenarios contribute to theta. We find that I-S3 cells suppress OLM cell spiking, rather than engender its spiking via post-inhibitory rebound mechanisms, and contribute to theta frequency spike resonance during simulated in vivo scenarios. To elicit recruitment similar to in vitro experiments, inclusion of disinhibited pyramidal cell inputs is necessary, implying that I-S3 cell firing broadens the window for pyramidal cell disinhibition. Using in vivo virtual networks, we show that I-S3 cells contribute to a sharpening of OLM cell recruitment at theta frequencies. Further, shifting the timing of I-S3 cell spiking due to external modulation shifts the timing of the OLM cell firing and thus disinhibitory windows. We propose a specialized contribution of I-S3 cells to create temporally precise coordination of modulation pathways.

Learning excitatory-inhibitory neuronal assemblies in recurrent networks

Understanding the connectivity observed in the brain and how it emerges from local plasticity rules is a grand challenge in modern neuroscience. In the primary visual cortex (V1) of mice, synapses between excitatory pyramidal neurons and inhibitory parvalbumin-expressing (PV) interneurons tend to be stronger for neurons that respond to similar stimulus features, although these neurons are not topographically arranged according to their stimulus preference. The presence of such excitatory-inhibitory (E/I) neuronal assemblies indicates a stimulus-specific form of feedback inhibition. Here, we show that activity-dependent synaptic plasticity on input and output synapses of PV interneurons generates a circuit structure that is consistent with mouse V1. Computational modeling reveals that both forms of plasticity must act in synergy to form the observed E/I assemblies. Once established, these assemblies produce a stimulus-specific competition between pyramidal neurons. Our model suggests that activity-dependent plasticity can refine inhibitory circuits to actively shape cortical computations.

Opposing Influence of Top-down and Bottom-up Input on Excitatory Layer 2/3 Neurons in Mouse Primary Visual Cortex

Processing in cortical circuits is driven by combinations of cortical and subcortical inputs. These inputs are often conceptually categorized as bottom-up, conveying sensory information, and top-down, conveying contextual information. Using intracellular recordings in mouse primary visual cortex, we measured neuronal responses to visual input, locomotion, and visuomotor mismatches. We show that layer 2/3 (L2/3) neurons compute a difference between top-down motor-related input and bottom-up visual flow input. Most L2/3 neurons responded to visuomotor mismatch with either hyperpolarization or depolarization, and the size of this response was correlated with distinct physiological properties. Consistent with a subtraction of bottom-up and top-down input, visual and motor-related inputs had opposing influence on L2/3 neurons. In infragranular neurons, we found no evidence of a difference computation and responses were consistent with positive integration of visuomotor inputs. Our results provide evidence that L2/3 functions as a bidirectional comparator of top-down and bottom-up input.

•

Layer 2/3 neurons show widespread subthreshold mismatch responses

•

Mismatch response sign is predicted by visual flow and locomotion-related responses

•

Layer 5/6 has a scarcity of depolarizing mismatch responses

•

Visual flow and locomotion speed have opposing signs of influence only in layer 2/3