One of the remarkable properties of cortical neurons is their innate ability to detect spike patterns in continuous but noisy information streams. As animals learn behaviorally relevant information quickly, pattern detection by neurons should also be rapid. However, how pattern-selective neuronal responses rapidly develop, tune, and remain robust has not been fully understood. Here, we propose a biologically plausible synaptic plasticity rule to learn patterned synaptic inputs rapidly. Our rule facilitates intracellular "self-supervised" learning of intermittently and repeatedly co-activated presynaptic-neuron communities, as in a machine-learning-based rule derived previously for temporal pattern segmentation. Importantly, this model proposes that a spike-triggered transient increase in somatodendritic coupling dramatically boosts the crediting of synapses responsible for the learned response. This boosting effect is essential for rapid pattern learning of single neurons at a high signal-to-noise ratio. Moreover, we demonstrate in recurrent networks how the rule recruits pre-configured cell assemblies for even faster, few-shot learning of multiple patterns. Our results illuminate the self-supervisory role of backpropagating action potentials for rapid pattern learning.

Spontaneous neural activity, crucial in memory, learning, and spatial navigation, often manifests itself as repetitive spatiotemporal patterns. Despite their importance, analyzing these patterns in large neural recordings remains challenging due to a lack of efficient and scalable detection methods. Addressing this gap, we introduce convSeq, an unsupervised method that employs backpropagation for optimizing spatiotemporal filters that effectively identify these neural patterns. Our method’s performance is validated on various synthetic data and real neural recordings, revealing spike sequences with unprecedented scalability and efficiency. Significantly surpassing existing methods in speed, convSeq sets a new standard for analyzing spontaneous neural activity, potentially advancing our understanding of information processing in neural circuits.

The agent learns to organize decision behavior to achieve a behavioral goal, such as reward maximization, and reinforcement learning is often used for this optimization. Learning an optimal behavioral strategy is difficult under the uncertainty that events necessary for learning are only partially observable, called as Partially Observable Markov Decision Process (POMDP). However, the real-world environment also gives many events irrelevant to reward delivery and an optimal behavioral strategy. The conventional methods in POMDP, which attempt to infer transition rules among the entire observations, including irrelevant states, are ineffective in such an environment. Supposing Redundantly Observable Markov Decision Process (ROMDP), here we propose a method for goal-oriented reinforcement learning to efficiently learn state transition rules among reward-related "core states" from redundant observations. Starting with a small number of initial core states, our model gradually adds new core states to the transition diagram until it achieves an optimal behavioral strategy consistent with the Bellman equation. We demonstrate that the resultant inference model outperforms the conventional method for POMDP. We emphasize that our model only containing the core states has high explainability. Furthermore, the proposed method suits online learning as it suppresses memory consumption and improves learning speed.

Advances in generative models have recently revolutionised machine learning. Meanwhile, in neuroscience, generative models have long been thought fundamental to animal intelligence. Understanding the biological mechanisms that support these processes promises to shed light on the relationship between biological and artificial intelligence. In animals, the hippocampal formation is thought to learn and use a generative model to support its role in spatial and non-spatial memory. Here we introduce a biologically plausible model of the hippocampal formation tantamount to a Helmholtz machine that we apply to a temporal stream of inputs. A novel component of our model is that fast theta-band oscillations (5-10 Hz) gate the direction of information flow throughout the network, training it akin to a high-frequency wake-sleep algorithm. Our model accurately infers the latent state of high-dimensional sensory environments and generates realistic sensory predictions. Furthermore, it can learn to path integrate by developing a ring attractor connectivity structure matching previous theoretical proposals and flexibly transfer this structure between environments. Whereas many models trade-off biological plausibility with generality, our model captures a variety of hippocampal cognitive functions under one biologically plausible local learning rule.

The brain performs various cognitive functions by learning the spatiotemporal salient features of the environment. This learning requires unsupervised segmentation of hierarchically organized spike sequences, but the underlying neural mechanism is only poorly understood. Here, we show that a recurrent gated network of neurons with dendrites can efficiently solve difficult segmentation tasks. In this model, multiplicative recurrent connections learn a context-dependent gating of dendro-somatic information transfers to minimize error in the prediction of somatic responses by the dendrites. Consequently, these connections filter the redundant input features represented by the dendrites but unnecessary in the given context. The model was tested on both synthetic and real neural data. In particular, the model was successful for segmenting multiple cell assemblies repeating in large-scale calcium imaging data containing thousands of cortical neurons. Our results suggest that recurrent gating of dendro-somatic signal transfers is crucial for cortical learning of context-dependent segmentation tasks.

Our cognition relies on the ability of the brain to segment hierarchically structured events on multiple scales. Recent evidence suggests that the brain performs this event segmentation based on the structure of state-transition graphs behind sequential experiences. However, the underlying circuit mechanisms are poorly understood. In this paper we propose an extended attractor network model for graph-based hierarchical computation which we call the Laplacian associative memory. This model generates multiscale representations for communities (clusters) of associative links between memory items, and the scale is regulated by the heterogenous modulation of inhibitory circuits. We analytically and numerically show that these representations correspond to graph Laplacian eigenvectors, a popular method for graph segmentation and dimensionality reduction. Finally, we demonstrate that our model exhibits chunked sequential activity patterns resembling hippocampal theta sequences. Our model connects graph theory and attractor dynamics to provide a biologically plausible mechanism for abstraction in the brain.

Chunking is the process by which frequently repeated segments of temporal inputs are concatenated into single units that are easy to process. Such a process is fundamental to time-series analysis in biological and artificial information processing systems. The brain efficiently acquires chunks from various information streams in an unsupervised manner; however, the underlying mechanisms of this process remain elusive. A widely-adopted statistical method for chunking consists of predicting frequently repeated contiguous elements in an input sequence based on unequal transition probabilities over sequence elements. However, recent experimental findings suggest that the brain is unlikely to adopt this method, as human subjects can chunk sequences with uniform transition probabilities. In this study, we propose a novel conceptual framework to overcome this limitation. In this process, neural networks learn to predict dynamical response patterns to sequence input rather than to directly learn transition patterns. Using a mutually supervising pair of reservoir computing modules, we demonstrate how this mechanism works in chunking sequences of letters or visual images with variable regularity and complexity. In addition, we demonstrate that background noise plays a crucial role in correctly learning chunks in this model. In particular, the model can successfully chunk sequences that conventional statistical approaches fail to chunk due to uniform transition probabilities. In addition, the neural responses of the model exhibit an interesting similarity to those of the basal ganglia observed after motor habit formation.

The brain can learn and detect mixed input signals masked by various types of noise, and spike-timing-dependent plasticity (STDP) is the candidate synaptic level mechanism. Because sensory inputs typically have spike correlation, and local circuits have dense feedback connections, input spikes cause the propagation of spike correlation in lateral circuits; however, it is largely unknown how this secondary correlation generated by lateral circuits influences learning processes through STDP, or whether it is beneficial to achieve efficient spike-based learning from uncertain stimuli. To explore the answers to these questions, we construct models of feedforward networks with lateral inhibitory circuits and study how propagated correlation influences STDP learning, and what kind of learning algorithm such circuits achieve. We derive analytical conditions at which neurons detect minor signals with STDP, and show that depending on the origin of the noise, different correlation timescales are useful for learning. In particular, we show that non-precise spike correlation is beneficial for learning in the presence of cross-talk noise. We also show that by considering excitatory and inhibitory STDP at lateral connections, the circuit can acquire a lateral structure optimal for signal detection. In addition, we demonstrate that the model performs blind source separation in a manner similar to the sequential sampling approximation of the Bayesian independent component analysis algorithm. Our results provide a basic understanding of STDP learning in feedback circuits by integrating analyses from both dynamical systems and information theory.

The connectivity structure of neural networks has significant implications for neural information processing, and much experimental effort has been made to clarify the structure of neural networks in the brain, i.e., both graph structure and weight structure of synaptic connections. A traditional view of neural information processing suggests that neurons compute in a highly parallel and distributed manner, in which the cooperation of many weak synaptic inputs is necessary to activate a single neuron. Recent experiments, however, have shown that not all synapses are weak in cortical circuits, but some synapses are extremely strong (several tens of times larger than the average weight). In fact, the weights of excitatory synapses between cortical excitatory neurons often obey a lognormal distribution with a long tail of strong synapses. Here, we review some of our important and recent works on computation with sparsely distributed synaptic weights and discuss the possible implications of this synaptic principle for neural computation by spiking neurons. We demonstrate that internal noise emerges from long-tailed distributions of synaptic weights to produce stochastic resonance effect in the reverberating synaptic pathways constituted by strong synapses. We show a spike-timing-dependent plasticity rule and other mechanisms that produce such weight distributions. A possible hardware realization of lognormally connected networks is also shown.

Spike-timing-dependent plasticity (STDP) has been observed in many brain areas such as sensory cortices, where it is hypothesized to structure synaptic connections between neurons. Previous studies have demonstrated how STDP can capture spiking information at short timescales using specific input configurations, such as coincident spiking, spike patterns and oscillatory spike trains. However, the corresponding computation in the case of arbitrary input signals is still unclear. This paper provides an overarching picture of the algorithm inherent to STDP, tying together many previous results for commonly used models of pairwise STDP. For a single neuron with plastic excitatory synapses, we show how STDP performs a spectral analysis on the temporal cross-correlograms between its afferent spike trains. The postsynaptic responses and STDP learning window determine kernel functions that specify how the neuron "sees" the input correlations. We thus denote this unsupervised learning scheme as 'kernel spectral component analysis' (kSCA). In particular, the whole input correlation structure must be considered since all plastic synapses compete with each other. We find that kSCA is enhanced when weight-dependent STDP induces gradual synaptic competition. For a spiking neuron with a "linear" response and pairwise STDP alone, we find that kSCA resembles principal component analysis (PCA). However, plain STDP does not isolate correlation sources in general, e.g., when they are mixed among the input spike trains. In other words, it does not perform independent component analysis (ICA). Tuning the neuron to a single correlation source can be achieved when STDP is paired with a homeostatic mechanism that reinforces the competition between synaptic inputs. Our results suggest that neuronal networks equipped with STDP can process signals encoded in the transient spiking activity at the timescales of tens of milliseconds for usual STDP.

The brain is considered to use a relatively small amount of energy for its efficient information processing. Under a severe restriction on the energy consumption, the maximization of mutual information (MMI), which is adequate for designing artificial processing machines, may not suit for the brain. The MMI attempts to send information as accurate as possible and this usually requires a sufficient energy supply for establishing clearly discretized communication bands. Here, we derive an alternative hypothesis for neural code from the neuronal activities recorded juxtacellularly in the sensorimotor cortex of behaving rats. Our hypothesis states that in vivo cortical neurons maximize the entropy of neuronal firing under two constraints, one limiting the energy consumption (as assumed previously) and one restricting the uncertainty in output spike sequences at given firing rate. Thus, the conditional maximization of firing-rate entropy (CMFE) solves a tradeoff between the energy cost and noise in neuronal response. In short, the CMFE sends a rich variety of information through broader communication bands (i.e., widely distributed firing rates) at the cost of accuracy. We demonstrate that the CMFE is reflected in the long-tailed, typically power law, distributions of inter-spike intervals obtained for the majority of recorded neurons. In other words, the power-law tails are more consistent with the CMFE rather than the MMI. Thus, we propose the mathematical principle by which cortical neurons may represent information about synaptic input into their output spike trains.

Temporal integration of input is essential to the accumulation of information in various cognitive and behavioral processes, and gradually increasing neuronal activity, typically occurring within a range of seconds, is considered to reflect such computation by the brain. Some psychological evidence suggests that temporal integration by the brain is nearly perfect, that is, the integration is non-leaky, and the output of a neural integrator is accurately proportional to the strength of input. Neural mechanisms of perfect temporal integration, however, remain largely unknown. Here, we propose a recurrent network model of cortical neurons that perfectly integrates partially correlated, irregular input spike trains. We demonstrate that the rate of this temporal integration changes proportionately to the probability of spike coincidences in synaptic inputs. We analytically prove that this highly accurate integration of synaptic inputs emerges from integration of the variance of the fluctuating synaptic inputs, when their mean component is kept constant. Highly irregular neuronal firing and spike coincidences are the major features of cortical activity, but they have been separately addressed so far. Our results suggest that the efficient protocol of information integration by cortical networks essentially requires both features and hence is heterotic.

The ability to make a correct choice of behavior from various options is crucial for animals' survival. The neural basis for the choice of behavior has been attracting growing attention in research on biological and artificial neural systems. Alternative choice tasks with variable ratio (VR) and variable interval (VI) schedules of reinforcement have often been employed in studying decision making by animals and humans. In the VR schedule task, alternative choices are reinforced with different probabilities, and subjects learn to select the behavioral response rewarded more frequently. In the VI schedule task, alternative choices are reinforced at different average intervals independent of the choice frequencies, and the choice behavior follows the so-called matching law. The two policies appear robustly in subjects' choice of behavior, but the underlying neural mechanisms remain unknown. Here, we show that these seemingly different policies can appear from a common computational algorithm known as actor-critic learning. We present experimentally testable variations of the VI schedule in which the matching behavior gives only a suboptimal solution to decision making and show that the actor-critic system exhibits the matching behavior in the steady state of the learning even when the matching behavior is suboptimal. However, it is found that the matching behavior can earn approximately the same reward as the optimal one in many practical situations.

Synaptic plasticity is considered to play a crucial role in the experience-dependent self-organization of local cortical networks. In the absence of sensory stimuli, cerebral cortex exhibits spontaneous membrane potential transitions between an UP and a DOWN state. To reveal how cortical networks develop spontaneous activity, or conversely, how spontaneous activity structures cortical networks, we analyze the self-organization of a recurrent network model of excitatory and inhibitory neurons, which is realistic enough to replicate UP-DOWN states, with spike-timing-dependent plasticity (STDP). The individual neurons in the self-organized network exhibit a variety of temporal patterns in the two-state transitions. In addition, the model develops a feed-forward network-like structure that produces a diverse repertoire of precise sequences of the UP state. Our model shows that the self-organized activity well resembles the spontaneous activity of cortical networks if STDP is accompanied by the pruning of weak synapses. These results suggest that the two-state membrane potential transitions play an active role in structuring local cortical circuits.

Much evidence indicates that synchronized gamma-frequency (20-70 Hz) oscillation plays a significant functional role in the neocortex and hippocampus. Chattering neuron is a possible neocortical pacemaker for the gamma oscillation. Based on our recent model of chattering neurons, here we study how gamma-frequency bursting is synchronized in a network of these neurons. Using a phase oscillator description, we first examine how two coupled chattering neurons are synchronized. The analysis reveals that an incremental change of the bursting mode, such as from singlet to doublet, always accompanies a rapid transition from antisynchronous to synchronous firing. The state transition occurs regardless of what changes the bursting mode. Within each bursting mode, the neuronal activity undergoes a gradual change from synchrony to antisynchrony. Since the sensitivity to Ca(2+) and the maximum conductance of Ca(2+)-dependent cationic current as well as the intensity of input current systematically control the bursting mode, these quantities may be crucial for the regulation of the coherence of local cortical activity. Numerical simulations demonstrate that the modulations of the calcium sensitivity and the amplitude of the cationic current can induce rapid transitions between synchrony and asynchrony in a large-scale network of chattering neurons. The rapid synchronization of chattering neurons is shown to synchronize the activities of regular spiking pyramidal neurons at the gamma frequencies, as may be necessary for selective attention or binding processing in object recognition.

Synapses in various neural preparations exhibit spike-timing-dependent plasticity (STDP) with a variety of learning window functions. The window functions determine the magnitude and the polarity of synaptic change according to the time difference of pre- and postsynaptic spikes. Numerical experiments revealed that STDP learning with a single-exponential window function resulted in a bimodal distribution of synaptic conductances as a consequence of competition between synapses. A slightly modified window function, however, resulted in a unimodal distribution rather than a bimodal distribution. Since various window functions have been observed in neural preparations, we develop a rigorous mathematical method to calculate the conductance distribution for any given window function. Our method is based on the Fokker-Planck equation to determine the conductance distribution and on the Ornstein-Uhlenbeck process to characterize the membrane potential fluctuations. Demonstrating that our method reproduces the known quantitative results of STDP learning, we apply the method to the type of STDP learning found recently in the CA1 region of the rat hippocampus. We find that this learning can result in nearly optimized competition between synapses. Meanwhile, we find that the type of STDP learning found in the cerebellum-like structure of electric fish can result in all-or-none synapses: either all the synaptic conductances are maximized, or none of them becomes significantly large. Our method also determines the window function that optimizes synaptic competition.

Fast-spiking (FS) interneurons have specific types (Kv3.1/3.2 type) of the delayed potassium channel, which differ from the conventional Hodgkin-Huxley (HH) type potassium channel (Kv1.3 type) in several aspects. In this study, we show dramatic effects of the Kv3.1/3.2 potassium channel on the synchronization of the FS interneurons. We show analytically that two identical electrically coupled FS interneurons modeled with Kv3.1/3.2 channel fire synchronously at arbitrary firing frequencies, unlike similarly coupled FS neurons modeled with Kv1.3 channel that show frequency-dependent synchronous and antisynchronous firing states. Introducing GABA(A) receptor-mediated synaptic connections into an FS neuron pair tends to induce an antisynchronous firing state, even if the chemical synapses are bidirectional. Accordingly, an FS neuron pair connected simultaneously by electrical and chemical synapses achieves both synchronous firing state and antisynchronous firing state in a physiologically plausible range of the conductance ratio between electrical and chemical synapses. Moreover, we find that a large-scale network of FS interneurons connected by gap junctions and bidirectional GABAergic synapses shows similar bistability in the range of gamma frequencies (30-70 Hz).

A neuroecological equation of the Lotka-Volterra type for mean firing rate is derived from the conventional membrane dynamics of a neural network with lateral inhibition and self-inhibition. Neural selection mechanisms employed by the competitive neural network receiving external inputs are studied with analytic and numerical calculations. A remarkable findings is that the strength of lateral inhibition relative to that of self-inhibition is crucial for determining the steady states of the network among three qualitatively different types of behavior. Equal strength of both types of inhibitory connections leads the network to the well-known winner-take-all behavior. If, however, the lateral inhibition is weaker than the self-inhibition, a certain number of neurons are activated in the steady states or the number of winners is in general more than one (the winners-share-all behavior). On the other hand, if the self-inhibition is weaker than the lateral one, only one neuron is activated, but the winner is not necessarily the neuron receiving the largest input. It is suggested that our simple network model provides a mathematical basis for understanding neural selection mechanisms.

An MOS circuit is proposed for implementing a nonmonotonic transfer characteristic of a neural network. The present research is motivated by the recent results of theoretical studies showing excellent equilibrium properties of networks with the nonmonotonic neural units. These properties include enhancement of storage capacity and complete elimination of noise in associative memory recall. The simple form of the transfer characteristic enables one to implement it with a simple electrical circuit of standard MOS transistors. SPICE simulation results are shown for the behavior of the neural units in associative memory recall.

It is shown that approximate fixed-point attractors rather than synchronized oscillations can be employed by a wide class of neural networks of oscillators to achieve an associative memory recall. This computational ability of oscillator neural networks is ensured by the fact that reduced dynamic equations for phase variables in general involve two terms that can be respectively responsible for the emergence of synchronization and cessation of oscillations. Thus the cessation occurs in memory retrieval if the corresponding term dominates in the dynamic equations. A bottomless feature of the energy function for such a system makes the retrieval states quasi-fixed points, which admit continual rotating motion to a small portion of oscillators, when an extensive number of memory patterns are embedded. An approximate theory based on the self-consistent signal-to-noise analysis enables one to study the equilibrium properties of the neural network of phase variables with the quasi-fixed-point attractors. As far as the memory retrieval by the quasi-fixed points is concerned, the equilibrium properties including the storage capacity of oscillator neural networks are proved to be similar to those of the Hopfield type neural networks.