Recurrent Connections Research Articles

Emergence of cortical network motifs for short-term memory during learning

Learning of adaptive behaviors requires the refinement of coordinated activity across multiple brain regions. However, how neural communications develop during learning remains poorly understood. Here, using two-photon calcium imaging, we simultaneously recorded the activity of layer 2/3 excitatory neurons in eight regions of the mouse dorsal cortex during learning of a delayed-response task. Across learning, while global functional connectivity became sparser, there emerged a subnetwork comprising of neurons in the anterior lateral motor cortex (ALM) and posterior parietal cortex (PPC). Neurons in this subnetwork shared a similar choice code during action preparation and formed recurrent functional connectivity across learning. Suppression of PPC activity disrupted choice selectivity in ALM and impaired task performance. Recurrent neural networks reconstructed from ALM activity revealed that PPC-ALM interactions rendered choice-related attractor dynamics more stable. Thus, learning constructs cortical network motifs by recruiting specific inter-areal communication channels to promote efficient and robust sensorimotor transformation.

Interaction Between HCN and Slack Channels Regulates mPFC Pyramidal Cell Excitability in Working Memory Circuits.

The ability of monkeys and rats to carry out spatial working memory tasks has been shown to depend on the persistent firing of pyramidal cells in the prefrontal cortex (PFC), arising from recurrent excitatory connections on dendritic spines. These spines express hyperpolarization-activated cyclic nucleotide-gated (HCN) channels whose open state is increased by cAMP signaling, and which markedly alter PFC network connectivity and neuronal firing. In traditional neural circuits, activation of these non-selective cation channels leads to neuronal depolarization and increased firing rate. Paradoxically, cAMP activation of HCN channels in PFC pyramidal cells reduces working memory-related neuronal firing. This suggests that activation of HCN channels may hyperpolarize rather than depolarize these neurons. The current study tested the hypothesis that Na+ influx through HCN channels activates Slack Na+-activated K+ (KNa) channels to hyperpolarize the membrane. We have found that HCN and Slack KNa channels co-immunoprecipitate in cortical extracts and that, by immunoelectron microscopy, they colocalize at postsynaptic spines of PFC pyramidal neurons. A specific blocker of HCN channels, ZD7288, reduces KNa current in pyramidal cells that express both HCN and Slack channels, but has no effect on KNa currents in an HEK cell line expressing Slack without HCN channels, indicating that blockade of HCN channels in neurons reduces K+ current indirectly by lowering Na+ influx. Activation of HCN channels by cAMP in a cell line expressing a Ca2+ reporter results in elevation of cytoplasmic Ca2+, but the effect of cAMP is reversed if the HCN channels are co-expressed with Slack channels. Finally, we used a novel pharmacological blocker of Slack channels to show that inhibition of Slack in rat PFC improves working memory performance, an effect previously demonstrated for blockers of HCN channels. Our results suggest that the regulation of working memory by HCN channels in PFC pyramidal neurons is mediated by an HCN-Slack channel complex that links activation HCN channels to suppression of neuronal excitability.

Excitation creates a distributed pattern of cortical suppression due to varied recurrent input

Dense local, recurrent connections are a major feature of cortical circuits, yet how they affect neurons' responses has been unclear, with some studies reporting weak recurrent effects, some reporting amplification, and others indicating local suppression. Here, we show that optogenetic input to mouse V1 excitatory neurons generates salt-and-pepper patterns of both excitation and suppression. Responses in individual neurons are not strongly predicted by that neuron's direct input. A balanced-state network model reconciles a set of diverse observations: the observed dynamics, suppressed responses, decoupling of input and output, and long tail of excited responses. The model shows recurrent excitatory-excitatory connections are strong and also variable across neurons. Together, these results demonstrate that excitatory recurrent connections can have major effects on cortical computations by shaping and changing neurons' responses to input.

Neural dynamic foundations of a theory of higher cognition: the case of grounding nested phrases.

Because cognitive competences emerge in evolution and development from the sensory-motor domain, we seek a neural process account for higher cognition in which all representations are necessarily grounded in perception and action. The challenge is to understand how hallmarks of higher cognition, productivity, systematicity, and compositionality, may emerge from such a bottom-up approach. To address this challenge, we present key ideas from Dynamic Field Theory which postulates that neural populations are organized by recurrent connectivity to create stable localist representations. Dynamic instabilities enable the autonomous generation of sequences of mental states. The capacity to apply neural circuitry across broad sets of inputs that emulates the function call postulated in symbolic computation emerges through coordinate transforms implemented in neural gain fields. We show how binding localist neural representations through a shared index dimension enables conceptual structure, in which the interdependence among components of a representation is flexibly expressed. We demonstrate these principles in a neural dynamic architecture that represents and perceptually grounds nested relational and action phrases. Sequences of neural processing steps are generated autonomously to attentionally select the referenced objects and events in a manner that is sensitive to their interdependencies. This solves the problem of 2 and the massive binding problem in expressions such as "the small tree that is to the left of the lake which is to the left of the large tree". We extend earlier work by incorporating new types of grammatical constructions and a larger vocabulary. We discuss the DFT framework relative to other neural process accounts of higher cognition and assess the scope and challenges of such neural theories.

Cluster-Based Input Weight Initialization for Echo State Networks.

Echo state networks (ESNs) are a special type of recurrent neural networks (RNNs), in which the input and recurrent connections are traditionally generated randomly, and only the output weights are trained. Despite the recent success of ESNs in various tasks of audio, image, and radar recognition, we postulate that a purely random initialization is not the ideal way of initializing ESNs. The aim of this work is to propose an unsupervised initialization of the input connections using the K-means algorithm on the training data. We show that for a large variety of datasets, this initialization performs equivalently or superior than a randomly initialized ESN while needing significantly less reservoir neurons. Furthermore, we discuss that this approach provides the opportunity to estimate a suitable size of the reservoir based on prior knowledge about the data.

A role for cortical interneurons as adversarial discriminators.

The brain learns representations of sensory information from experience, but the algorithms by which it does so remain unknown. One popular theory formalizes representations as inferred factors in a generative model of sensory stimuli, meaning that learning must improve this generative model and inference procedure. This framework underlies many classic computational theories of sensory learning, such as Boltzmann machines, the Wake/Sleep algorithm, and a more recent proposal that the brain learns with an adversarial algorithm that compares waking and dreaming activity. However, in order for such theories to provide insights into the cellular mechanisms of sensory learning, they must be first linked to the cell types in the brain that mediate them. In this study, we examine whether a subtype of cortical interneurons might mediate sensory learning by serving as discriminators, a crucial component in an adversarial algorithm for representation learning. We describe how such interneurons would be characterized by a plasticity rule that switches from Hebbian plasticity during waking states to anti-Hebbian plasticity in dreaming states. Evaluating the computational advantages and disadvantages of this algorithm, we find that it excels at learning representations in networks with recurrent connections but scales poorly with network size. This limitation can be partially addressed if the network also oscillates between evoked activity and generative samples on faster timescales. Consequently, we propose that an adversarial algorithm with interneurons as discriminators is a plausible and testable strategy for sensory learning in biological systems.

Exploring unsupervised pre-training for echo state networks

Echo State Networks (ESNs) are a special type of Recurrent Neural Networks (RNNs), in which the input and recurrent connections are traditionally generated randomly, and only the output weights are trained. However, recent publications have addressed the problem that a purely random initialization may not be ideal. Instead, a completely deterministic or data-driven initialized ESN structure was proposed. In this work, an unsupervised training methodology for the hidden components of an ESN is proposed. Motivated by traditional Hidden Markov Models (HMMs), which have been widely used for speech recognition for decades, we present an unsupervised pre-training method for the recurrent weights and bias weights of ESNs. This approach allows for using unlabeled data during the training procedure and shows superior results for continuous spoken phoneme recognition, as well as for a large variety of time-series classification datasets.

A 22-pJ/spike 73-Mspikes/s 130k-compartment neural array transceiver with conductance-based synaptic and membrane dynamics.

Neuromorphic cognitive computing offers a bio-inspired means to approach the natural intelligence of biological neural systems in silicon integrated circuits. Typically, such circuits either reproduce biophysical neuronal dynamics in great detail as tools for computational neuroscience, or abstract away the biology by simplifying the functional forms of neural computation in large-scale systems for machine intelligence with high integration density and energy efficiency. Here we report a hybrid which offers biophysical realism in the emulation of multi-compartmental neuronal network dynamics at very large scale with high implementation efficiency, and yet with high flexibility in configuring the functional form and the network topology. The integrate-and-fire array transceiver (IFAT) chip emulates the continuous-time analog membrane dynamics of 65 k two-compartment neurons with conductance-based synapses. Fired action potentials are registered as address-event encoded output spikes, while the four types of synapses coupling to each neuron are activated by address-event decoded input spikes for fully reconfigurable synaptic connectivity, facilitating virtual wiring as implemented by routing address-event spikes externally through synaptic routing table. Peak conductance strength of synapse activation specified by the address-event input spans three decades of dynamic range, digitally controlled by pulse width and amplitude modulation (PWAM) of the drive voltage activating the log-domain linear synapse circuit. Two nested levels of micro-pipelining in the IFAT architecture improve both throughput and efficiency of synaptic input. This two-tier micro-pipelining results in a measured sustained peak throughput of 73 Mspikes/s and overall chip-level energy efficiency of 22 pJ/spike. Non-uniformity in digitally encoded synapse strength due to analog mismatch is mitigated through single-point digital offset calibration. Combined with the flexibly layered and recurrent synaptic connectivity provided by hierarchical address-event routing of registered spike events through external memory, the IFAT lends itself to efficient large-scale emulation of general biophysical spiking neural networks, as well as rate-based mapping of rectified linear unit (ReLU) neural activations.

Explainable district heat load forecasting with active deep learning

District heat load forecasting is a challenging task that involves predicting future heat demand based on historical data and various influencing factors. Accurate forecasting is essential for optimizing energy production and distribution in district heating systems. However, most existing forecasting models lack transparency and interpretability and fail to capture the spatial–temporal dependencies in the data. Moreover, they often require a large amount of annotated data for training, which can be costly and time-consuming to obtain. In this paper, we present a novel approach to district heat load forecasting, which involves predicting future heat demand based on historical data and various influencing factors. The proposed approach is based on an Active Graph Recurrent Network (Ac-GRN), which leverages the strengths of active deep learning and graph neural networks to capture the complex spatial–temporal dependencies in the data. The approach also provides explainability for its predictions by using correlation-based attribution methods. The active deep learning component can effectively select the most informative and representative samples from a large pool of data, reducing the frequency and cost of data collection and human effort. The graph neural network component can model both linear and nonlinear relationships among heat meters using bidirectional recurrent connections, enhancing the accuracy and robustness of the predictions. We conduct extensive experiments and compare our approach with eleven state-of-the-art models on a real-world dataset of district heating consumption data from Danish residential buildings. Our results show that our approach outperforms other models in terms of accuracy, robustness, reliability, and computational efficiency for multi-horizon multi-step district heat load forecasting. Our approach also provides meaningful explanations for its predictions by highlighting the most influential factors and heat meters for each prediction. This paper makes a novel contribution to district heat load forecasting with explainability.

Bio-inspired, task-free continual learning through activity regularization.

The ability to sequentially learn multiple tasks without forgetting is a key skill of biological brains, whereas it represents a major challenge to the field of deep learning. To avoid catastrophic forgetting, various continual learning (CL) approaches have been devised. However, these usually require discrete task boundaries. This requirement seems biologically implausible and often limits the application of CL methods in the real world where tasks are not always well defined. Here, we take inspiration from neuroscience, where sparse, non-overlapping neuronal representations have been suggested to prevent catastrophic forgetting. As in the brain, we argue that these sparse representations should be chosen on the basis of feed forward (stimulus-specific) as well as top-down (context-specific) information. To implement such selective sparsity, we use a bio-plausible form of hierarchical credit assignment known as Deep Feedback Control (DFC) and combine it with a winner-take-all sparsity mechanism. In addition to sparsity, we introduce lateral recurrent connections within each layer to further protect previously learned representations. We evaluate the new sparse-recurrent version of DFC on the split-MNIST computer vision benchmark and show that only the combination of sparsity and intra-layer recurrent connections improves CL performance with respect to standard backpropagation. Our method achieves similar performance to well-known CL methods, such as Elastic Weight Consolidation and Synaptic Intelligence, without requiring information about task boundaries. Overall, we showcase the idea of adopting computational principles from the brain to derive new, task-free learning algorithms for CL.

Warming up recurrent neural networks to maximise reachable multistability greatly improves learning

Training recurrent neural networks is known to be difficult when time dependencies become long. In this work, we show that most standard cells only have one stable equilibrium at initialisation, and that learning on tasks with long time dependencies generally occurs once the number of network stable equilibria increases; a property known as multistability. Multistability is often not easily attained by initially monostable networks, making learning of long time dependencies between inputs and outputs difficult. This insight leads to the design of a novel way to initialise any recurrent cell connectivity through a procedure called “warmup” to improve its capability to learn arbitrarily long time dependencies. This initialisation procedure is designed to maximise network reachable multistability, i.e., the number of equilibria within the network that can be reached through relevant input trajectories, in few gradient steps. We show on several information restitution, sequence classification, and reinforcement learning benchmarks that warming up greatly improves learning speed and performance, for multiple recurrent cells, but sometimes impedes precision. We therefore introduce a double-layer architecture initialised with a partial warmup that is shown to greatly improve learning of long time dependencies while maintaining high levels of precision. This approach provides a general framework for improving learning abilities of any recurrent cell when long time dependencies are present. We also show empirically that other initialisation and pretraining procedures from the literature implicitly foster reachable multistability of recurrent cells.

Top-down and within-layer recurrent connections in artificial networks are needed to solve challenging visual tasks

Top-down and within-layer recurrent connections in artificial networks are needed to solve challenging visual tasks

Dynamic Recruitment of the Feedforward and Recurrent Mechanism for Black-White Asymmetry in the Primary Visual Cortex.

Black and white information is asymmetrically distributed in natural scenes, evokes asymmetric neuronal responses, and causes asymmetric perceptions. Recognizing the universality and essentiality of black-white asymmetry in visual information processing, the neural substrates for black-white asymmetry remain unclear. To disentangle the role of the feedforward and recurrent mechanisms in the generation of cortical black-white asymmetry, we recorded the V1 laminar responses and LGN responses of anesthetized cats of both sexes. In a cortical column, we found that black-white asymmetry starts at the input layer and becomes more pronounced in the output layer. We also found distinct dynamics of black-white asymmetry between the output layer and the input layer. Specifically, black responses dominate in all layers after stimulus onset. After stimulus offset, black and white responses are balanced in the input layer, but black responses still dominate in the output layer. Compared with that in the input layer, the rebound response in the output layer is significantly suppressed. The relative suppression strength evoked by white stimuli is notably stronger and depends on the location within the ON-OFF cortical map. A model with delayed and polarity-selective cortical suppression explains black-white asymmetry in the output layer, within which prominent recurrent connections are identified by Granger causality analysis. In addition to black-white asymmetry in response strength, the interlaminar differences in spatial receptive field varied dynamically. Our findings suggest that the feedforward and recurrent mechanisms are dynamically recruited for the generation of black-white asymmetry in V1.SIGNIFICANCE STATEMENT Black-white asymmetry is universal and essential in visual information processing, yet the neural substrates for cortical black-white asymmetry remain unknown. Leveraging V1 laminar recordings, we provided the first laminar pattern of black-white asymmetry in cat V1 and found distinct dynamics of black-white asymmetry between the output layer and the input layer. Comparing black-white asymmetry across three visual hierarchies, the LGN, V1 input layer, and V1 output layer, we demonstrated that the feedforward and recurrent mechanisms are dynamically recruited for the generation of cortical black-white asymmetry. Our findings not only enhance our understanding of laminar processing within a cortical column but also elucidate how feedforward connections and recurrent connections interact to shape neuronal response properties.

Uncovering circuit mechanisms of current sinks and sources with biophysical simulations of primary visual cortex.

Local field potential (LFP) recordings reflect the dynamics of the current source density (CSD) in brain tissue. The synaptic, cellular, and circuit contributions to current sinks and sources are ill-understood. We investigated these in mouse primary visual cortex using public Neuropixels recordings and a detailed circuit model based on simulating the Hodgkin-Huxley dynamics of >50,000 neurons belonging to 17 cell types. The model simultaneously captured spiking and CSD responses and demonstrated a two-way dissociation: firing rates are altered with minor effects on the CSD pattern by adjusting synaptic weights, and CSD is altered with minor effects on firing rates by adjusting synaptic placement on the dendrites. We describe how thalamocortical inputs and recurrent connections sculpt specific sinks and sources early in the visual response, whereas cortical feedback crucially alters them in later stages. These results establish quantitative links between macroscopic brain measurements (LFP/CSD) and microscopic biophysics-based understanding of neuron dynamics and show that CSD analysis provides powerful constraints for modeling beyond those from considering spikes.

Inhibitory stabilized network behaviour in a balanced neural mass model of a cortical column

Strong inhibitory recurrent connections can reduce the tendency for a neural network to become unstable. This is known as inhibitory stabilization; networks that are unstable in the absence of strong inhibitory feedback because of their unstable excitatory recurrent connections are known as Inhibition Stabilized Networks (ISNs). One of the characteristics of ISNs is their “paradoxical response”, where perturbing the inhibitory neurons with additional excitatory input results in a decrease in their activity after a temporal delay instead of increasing their activity. Here, we develop a model of populations of neurons across different layers of cortex. Within each layer, there is one population of inhibitory neurons and one population of excitatory neurons. The connectivity weights across different populations in the model are derived from a synaptic physiology database provided by the Allen Institute. The model shows a gradient of excitation–inhibition balance across different layers in the cortex, where superficial layers are more inhibitory dominated compared to deeper layers. To investigate the presence of ISNs across different layers, we measured the membrane potentials of neural populations in the model after perturbing inhibitory populations. The results show that layer 2/3 in the model does not operate in the ISN regime but layers 4 and 5 do operate in the ISN regime. These results accord with neurophysiological findings that explored the presence of ISNs across different layers in the cortex. The results show that there may be a systematic macroscopic gradient of inhibitory stabilization across different layers in the cortex that depends on the level of excitation–inhibition balance, and that the strength of the paradoxical response increases as the model moves closer to bifurcation points.

Cortical recurrence supports resilience to sensory variance in the primary visual cortex

Our daily endeavors occur in a complex visual environment, whose intrinsic variability challenges the way we integrate information to make decisions. By processing myriads of parallel sensory inputs, our brain is theoretically able to compute the variance of its environment, a cue known to guide our behavior. Yet, the neurobiological and computational basis of such variance computations are still poorly understood. Here, we quantify the dynamics of sensory variance modulations of cat primary visual cortex neurons. We report two archetypal neuronal responses, one of which is resilient to changes in variance and co-encodes the sensory feature and its variance, improving the population encoding of orientation. The existence of these variance-specific responses can be accounted for by a model of intracortical recurrent connectivity. We thus propose that local recurrent circuits process uncertainty as a generic computation, advancing our understanding of how the brain handles naturalistic inputs.

Waves traveling over a map of visual space can ignite short-term predictions of sensory input

Recent analyses have found waves of neural activity traveling across entire visual cortical areas in awake animals. These traveling waves modulate the excitability of local networks and perceptual sensitivity. The general computational role of these spatiotemporal patterns in the visual system, however, remains unclear. Here, we hypothesize that traveling waves endow the visual system with the capacity to predict complex and naturalistic inputs. We present a network model whose connections can be rapidly and efficiently trained to predict individual natural movies. After training, a few input frames from a movie trigger complex wave patterns that drive accurate predictions many frames into the future solely from the network’s connections. When the recurrent connections that drive waves are randomly shuffled, both traveling waves and the ability to predict are eliminated. These results suggest traveling waves may play an essential computational role in the visual system by embedding continuous spatiotemporal structures over spatial maps.

Neurochemistry-enriched dynamic causal models of magnetoencephalography, using magnetic resonance spectroscopy

We present a hierarchical empirical Bayesian framework for testing hypotheses about neurotransmitters’ concertation as empirical prior for synaptic physiology using ultra-high field magnetic resonance spectroscopy (7T-MRS) and magnetoencephalography data (MEG). A first level dynamic causal modelling of cortical microcircuits is used to infer the connectivity parameters of a generative model of individuals’ neurophysiological observations. At the second level, individuals’ 7T-MRS estimates of regional neurotransmitter concentration supply empirical priors on synaptic connectivity. We compare the group-wise evidence for alternative empirical priors, defined by monotonic functions of spectroscopic estimates, on subsets of synaptic connections. For efficiency and reproducibility, we used Bayesian model reduction (BMR), parametric empirical Bayes and variational Bayesian inversion. In particular, we used Bayesian model reduction to compare alternative model evidence of how spectroscopic neurotransmitter measures inform estimates of synaptic connectivity. This identifies the subset of synaptic connections that are influenced by individual differences in neurotransmitter levels, as measured by 7T-MRS. We demonstrate the method using resting-state MEG (i.e., task-free recording) and 7T-MRS data from healthy adults. Our results confirm the hypotheses that GABA concentration influences local recurrent inhibitory intrinsic connectivity in deep and superficial cortical layers, while glutamate influences the excitatory connections between superficial and deep layers and connections from superficial to inhibitory interneurons. Using within-subject split-sampling of the MEG dataset (i.e., validation by means of a held-out dataset), we show that model comparison for hypothesis testing can be highly reliable. The method is suitable for applications with magnetoencephalography or electroencephalography, and is well-suited to reveal the mechanisms of neurological and psychiatric disorders, including responses to psychopharmacological interventions.

Cyclic structure with cellular precision in a vertebrate sensorimotor neural circuit

Neuronal wiring diagrams reconstructed by electron microscopy1,2,3,4,5 pose new questions about the organization of nervous systems following the time-honored tradition of cross-species comparisons.6,7 The C.elegans connectome has been conceptualized as a sensorimotor circuit that is approximately feedforward,8,9,10,11 starting from sensory neurons proceeding to interneurons and ending with motor neurons. Overrepresentation of a 3-cell motif often known as the "feedforward loop" has provided further evidence for feedforwardness.10,12 Here, we contrast with another sensorimotor wiring diagram that was recently reconstructed from a larval zebrafish brainstem.13 We show that the 3-cycle, another 3-cell motif, is highly overrepresented in the oculomotor module of this wiring diagram. This is a first for any neuronal wiring diagram reconstructed by electron microscopy, whether invertebrate12,14 or mammalian.15,16,17 The 3-cycle of cells is "aligned" with a 3-cycle of neuronal groups in a stochastic block model (SBM)18 of the oculomotor module. However, the cellular cycles exhibit more specificity than can be explained by the group cycles-recurrence to the same neuron is surprisingly common. Cyclic structure could be relevant for theories of oculomotor function that depend on recurrent connectivity. The cyclic structure coexists with the classic vestibulo-ocular reflex arc for horizontal eye movements,19 and could be relevant for recurrent network models of temporal integration by the oculomotor system.20,21.

Self-stabilizing systems in spite of high dynamics

We initiate research on self-stabilization in highly dynamic identified message-passing systems where dynamics is modeled using time-varying graphs (TVGs). More precisely, we address the self-stabilizing leader election problem in three wide classes of TVGs: the class TCB(Δ) of TVGs with temporal diameter bounded by Δ, the class TCQ(Δ) of TVGs with temporal diameter quasi-bounded by Δ, and the class TCR of TVGs with recurrent connectivity only, where TCB(Δ)⊆TCQ(Δ)⊆TCR. We first study conditions under which our problem can be solved. We introduce the notion of size-ambiguity to show that the assumption on the knowledge of the number n of processes is central. Our results reveal that, despite the existence of unique process identifiers, any deterministic self-stabilizing leader election algorithm working on the class TCQ(Δ) or TCR cannot be size-ambiguous, justifying why our solutions for those classes assume the exact knowledge of n. We then present three self-stabilizing leader election algorithms for Classes TCB(Δ), TCQ(Δ), and TCR, respectively. Our algorithm for TCB(Δ) stabilizes in at most 3Δ rounds. However, we show that stabilization time cannot be bounded for the leader election problem in TCQ(Δ) and TCR. Nevertheless, we circumvent this issue by showing that our solutions are speculative in the sense that their stabilization time in TCB(Δ) (⊆TCQ(Δ)⊆TCR) is O(Δ) rounds.