Learning with three factors: modulating Hebbian plasticity with errors

Learning sequence attractors in recurrent networks with hidden neurons

The brain is targeted for processing temporal sequence information. It remains largely unclear how the brain learns to store and retrieve sequence memories. Here, we study how recurrent networks of binary neurons learn sequence attractors to store predefined pattern sequences and retrieve them robustly. We show that to store arbitrary pattern sequences, it is necessary for the network to include hidden neurons even though their role in displaying sequence memories is indirect. We develop a local learning algorithm to learn sequence attractors in the networks with hidden neurons. The algorithm is proven to converge and lead to sequence attractors. We demonstrate that the network model can store and retrieve sequences robustly on synthetic and real-world datasets. We hope that this study provides new insights in understanding sequence memory and temporal information processing in the brain.

Self-Induced Light Emission in Solid-State Memristors Replicates Neuronal Biophotons

Key neuronal functions have been successfully replicated in various hardware systems. Noticeable examples are neuronal networks constructed from memristors, which emulate complex electrochemical biological dynamics such as the efficacy and plasticity of a neuron. Neurons are highly active cells, communicating with chemical and electrical stimuli, but also emit light. These so-called biophotons are suspected to be a complementary vehicle to transport information across the brain. Here, we show that a memristor also releases photons during its operation akin to the production of neuronal light. Critical attributes of biophotons, such as self-generation, stochasticity, spectral coverage, sparsity, and correlation with the neuron's electrical activity, are replicated by our solid-state approach. Importantly, our time-resolved analysis of the correlated current transport and photon activity shows that emission takes place within a nanometer-sized active area and relies on electrically induced single-to-few active electroluminescent centers excited with moderate voltage (<3 V). Our findings further extend the emulating capability of a memristor to encompass neuronal optical activity and allow to construct memristive atomic-scale devices capable of handling simultaneously electrons and photons as information carriers.

Synaptic weight dynamics underlying memory consolidation: implications for learning rules, circuit organization, and circuit function

Systems consolidation is a common feature of learning and memory systems, in which a long-term memory initially stored in one brain region becomes persistently stored in another region. We studied the dynamics of systems consolidation in simple circuit architectures with two sites of plasticity, one in an early-learning and one in a late-learning brain area. We show that the synaptic dynamics of the circuit during consolidation of an analog memory can be understood as a temporal integration process, by which transient changes in activity driven by plasticity in the early-learning area are accumulated into persistent synaptic changes at the late-learning site. This simple principle naturally leads to a speed-accuracy tradeoff in systems consolidation and provides insight into how the circuit mitigates the stability-plasticity dilemma of storing new memories while preserving core features of older ones. Furthermore, it imposes two constraints on the circuit. First, the plasticity rule at the late-learning site must stably support a continuum of possible outputs for a given input. We show that this is readily achieved by heterosynaptic but not standard Hebbian rules. Second, to turn off the consolidation process and prevent erroneous changes at the late-learning site, neural activity in the early-learning area must be reset to its baseline activity. We propose two biologically plausible implementations for this reset that suggest novel roles for core elements of the cerebellar circuit.

Significance Statement

How are memories transformed over time? We propose a simple organizing principle for how long term memories are moved from an initial to a final site of storage. We show that successful transfer occurs when the late site of memory storage is endowed with synaptic plasticity rules that stably accumulate changes in activity occurring at the early site of memory storage. We instantiate this principle in a simple computational model that is representative of brain circuits underlying a variety of behaviors. The model suggests how a neural circuit can store new memories while preserving core features of older ones, and suggests novel roles for core elements of the cerebellar circuit.

Anti-Hebbian plasticity drives sequence learning in striatum

Spatio-temporal activity patterns have been observed in a variety of brain areas in spontaneous activity, prior to or during action, or in response to stimuli. Biological mechanisms endowing neurons with the ability to distinguish between different sequences remain largely unknown. Learning sequences of spikes raises multiple challenges, such as maintaining in memory spike history and discriminating partially overlapping sequences. Here, we show that anti-Hebbian spike-timing dependent plasticity (STDP), as observed at cortico-striatal synapses, can naturally lead to learning spike sequences. We design a spiking model of the striatal output neuron receiving spike patterns defined as sequential input from a fixed set of cortical neurons. We use a simple synaptic plasticity rule that combines anti-Hebbian STDP and non-associative potentiation for a subset of the presented patterns called rewarded patterns. We study the ability of striatal output neurons to discriminate rewarded from non-rewarded patterns by firing only after the presentation of a rewarded pattern. In particular, we show that two biological properties of striatal networks, spiking latency and collateral inhibition, contribute to an increase in accuracy, by allowing a better discrimination of partially overlapping sequences. These results suggest that anti-Hebbian STDP may serve as a biological substrate for learning sequences of spikes.

Chaotic neural dynamics facilitate probabilistic computations through sampling

Cortical neurons exhibit highly variable responses over trials and time. Theoretical works posit that this variability arises potentially from chaotic network dynamics of recurrently connected neurons. Here, we demonstrate that chaotic neural dynamics, formed through synaptic learning, allow networks to perform sensory cue integration in a sampling-based implementation. We show that the emergent chaotic dynamics provide neural substrates for generating samples not only of a static variable but also of a dynamical trajectory, where generic recurrent networks acquire these abilities with a biologically plausible learning rule through trial and error. Furthermore, the networks generalize their experience in the stimulus-evoked samples to the inference without partial or all sensory information, which suggests a computational role of spontaneous activity as a representation of the priors as well as a tractable biological computation for marginal distributions. These findings suggest that chaotic neural dynamics may serve for the brain function as a Bayesian generative model.

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.

Plasticity-stability dynamics during post-training processing of learning

Learning continues beyond the end of training. Post-training learning is supported by changes in plasticity and stability in the brain during both wakefulness and sleep. However, the lack of a unified measure for assessing plasticity and stability dynamics during training and post-training periods has limited our understanding of how these dynamics shape learning. Focusing primarily on procedural learning, we integrate work using behavioral paradigms and a recently developed measure, the excitatory-to-inhibitory (E/I) ratio, to explore the delicate balance between plasticity and stability and its relationship to post-training learning. This reveals plasticity-stability cycles during both wakefulness and sleep that enhance learning and protect it from new learning during post-training processing.

The combination of Hebbian and predictive plasticity learns invariant object representations in deep sensory networks

Recognition of objects from sensory stimuli is essential for survival. To that end, sensory networks in the brain must form object representations invariant to stimulus changes, such as size, orientation and context. Although Hebbian plasticity is known to shape sensory networks, it fails to create invariant object representations in computational models, raising the question of how the brain achieves such processing. In the present study, we show that combining Hebbian plasticity with a predictive form of plasticity leads to invariant representations in deep neural network models. We derive a local learning rule that generalizes to spiking neural networks and naturally accounts for several experimentally observed properties of synaptic plasticity, including metaplasticity and spike-timing-dependent plasticity. Finally, our model accurately captures neuronal selectivity changes observed in the primate inferotemporal cortex in response to altered visual experience. Thus, we provide a plausible normative theory emphasizing the importance of predictive plasticity mechanisms for successful representational learning.

Self-Organization of Nonlinearly Coupled Neural Fluctuations Into Synergistic Population Codes

Neural activity in the brain exhibits correlated fluctuations that may strongly influence the properties of neural population coding. However, how such correlated neural fluctuations may arise from the intrinsic neural circuit dynamics and subsequently affect the computational properties of neural population activity remains poorly understood. The main difficulty lies in resolving the nonlinear coupling between correlated fluctuations with the overall dynamics of the system. In this study, we investigate the emergence of synergistic neural population codes from the intrinsic dynamics of correlated neural fluctuations in a neural circuit model capturing realistic nonlinear noise coupling of spiking neurons. We show that a rich repertoire of spatial correlation patterns naturally emerges in a bump attractor network and further reveals the dynamical regime under which the interplay between differential and noise correlations leads to synergistic codes. Moreover, we find that negative correlations may induce stable bound states between two bumps, a phenomenon previously unobserved in firing rate models. These noise-induced effects of bump attractors lead to a number of computational advantages including enhanced working memory capacity and efficient spatiotemporal multiplexing and can account for a range of cognitive and behavioral phenomena related to working memory. This study offers a dynamical approach to investigating realistic correlated neural fluctuations and insights to their roles in cortical computations.

Canonical circuit computations for computer vision

Advanced computer vision mechanisms have been inspired by neuroscientific findings. However, with the focus on improving benchmark achievements, technical solutions have been shaped by application and engineering constraints. This includes the training of neural networks which led to the development of feature detectors optimally suited to the application domain. However, the limitations of such approaches motivate the need to identify computational principles, or motifs, in biological vision that can enable further foundational advances in machine vision. We propose to utilize structural and functional principles of neural systems that have been largely overlooked. They potentially provide new inspirations for computer vision mechanisms and models. Recurrent feedforward, lateral, and feedback interactions characterize general principles underlying processing in mammals. We derive a formal specification of core computational motifs that utilize these principles. These are combined to define model mechanisms for visual shape and motion processing. We demonstrate how such a framework can be adopted to run on neuromorphic brain-inspired hardware platforms and can be extended to automatically adapt to environment statistics. We argue that the identified principles and their formalization inspires sophisticated computational mechanisms with improved explanatory scope. These and other elaborated, biologically inspired models can be employed to design computer vision solutions for different tasks and they can be used to advance neural network architectures of learning.

The study of plasticity has always been about gradients

The experimental study of learning and plasticity has always been driven by an implicit question: how can physiological changes be adaptive and improve performance? For example, in Hebbian plasticity only synapses from presynaptic neurons that were active are changed, avoiding useless changes. Similarly, in dopamine-gated learning synapse changes depend on reward or lack thereof and do not change when everything is predictable. Within machine learning we can make the question of which changes are adaptive concrete: performance improves when changes correlate with the gradient of an objective function quantifying performance. This result is general for any system that improves through small changes. As such, physiology has always implicitly been seeking mechanisms that allow the brain to approximate gradients. Coming from this perspective we review the existing literature on plasticity-related mechanisms, and we show how these mechanisms relate to gradient estimation. We argue that gradients are a unifying idea to explain the many facets of neuronal plasticity.

Distinctive properties of biological neural networks and recent advances in bottom-up approaches toward a better biologically plausible neural network

Although it may appear infeasible and impractical, building artificial intelligence (AI) using a bottom-up approach based on the understanding of neuroscience is straightforward. The lack of a generalized governing principle for biological neural networks (BNNs) forces us to address this problem by converting piecemeal information on the diverse features of neurons, synapses, and neural circuits into AI. In this review, we described recent attempts to build a biologically plausible neural network by following neuroscientifically similar strategies of neural network optimization or by implanting the outcome of the optimization, such as the properties of single computational units and the characteristics of the network architecture. In addition, we proposed a formalism of the relationship between the set of objectives that neural networks attempt to achieve, and neural network classes categorized by how closely their architectural features resemble those of BNN. This formalism is expected to define the potential roles of top-down and bottom-up approaches for building a biologically plausible neural network and offer a map helping the navigation of the gap between neuroscience and AI engineering.

An integrated measure of GABA to characterize post-stroke plasticity

Stroke is a major cause of death and chronic neurological disability. Despite the improvements in stroke care, the number of patients affected by stroke keeps increasing and many stroke survivors are left permanently disabled. Current therapies are limited in efficacy. Understanding the neurobiological mechanisms underlying post-stroke recovery is therefore crucial to find new therapeutic options to address this medical burden. Long-lasting and widespread alterations of γ-aminobutyric acid (GABA) neurotransmission seem to play a key role in stroke recovery. In this review we first discuss a possible model of GABAergic modulation of post-stroke plasticity. We then overview the techniques currently available to non-invasively assess GABA in patients and the conclusions drawn from this limited body of work. Finally, we address the remaining open questions to clarify GABAergic changes underlying post-stroke recovery, we briefly review possible ways to modulate GABA post stroke and propose a novel approach to thoroughly quantify GABA in stroke patients, by integrating its concentration, the activity of its receptors and its link with microstructural changes.

Computational models of episodic-like memory in food-caching birds

Birds of the crow family adapt food-caching strategies to anticipated needs at the time of cache recovery and rely on memory of the what, where and when of previous caching events to recover their hidden food. It is unclear if this behavior can be explained by simple associative learning or if it relies on higher cognitive processes like mental time-travel. We present a computational model and propose a neural implementation of food-caching behavior. The model has hunger variables for motivational control, reward-modulated update of retrieval and caching policies and an associative neural network for remembering caching events with a memory consolidation mechanism for flexible decoding of the age of a memory. Our methodology of formalizing experimental protocols is transferable to other domains and facilitates model evaluation and experiment design. Here, we show that memory-augmented, associative reinforcement learning without mental time-travel is sufficient to explain the results of 28 behavioral experiments with food-caching birds.

Meta-SpikePropamine: learning to learn with synaptic plasticity in spiking neural networks

We propose that in order to harness our understanding of neuroscience toward machine learning, we must first have powerful tools for training brain-like models of learning. Although substantial progress has been made toward understanding the dynamics of learning in the brain, neuroscience-derived models of learning have yet to demonstrate the same performance capabilities as methods in deep learning such as gradient descent. Inspired by the successes of machine learning using gradient descent, we introduce a bi-level optimization framework that seeks to both solve online learning tasks and improve the ability to learn online using models of plasticity from neuroscience. We demonstrate that models of three-factor learning with synaptic plasticity taken from the neuroscience literature can be trained in Spiking Neural Networks (SNNs) with gradient descent via a framework of learning-to-learn to address challenging online learning problems. This framework opens a new path toward developing neuroscience inspired online learning algorithms.

Cellular bases for reward-related dopamine actions

Dopamine neurons exhibit transient increases and decreases in their firing rate upon reward and punishment for learning. This bidirectional modulation of dopamine dynamics occurs on the order of hundreds of milliseconds, and it is sensitively detected to reinforce the preceding sensorimotor events. These observations indicate that the mechanisms of dopamine detection at the projection sites are of remarkable precision, both in time and concentration. A major target of dopamine projection is the striatum, including the ventral region of the nucleus accumbens, which mainly comprises dopamine D1 and D2 receptor (D1R and D2R)-expressing spiny projection neurons. Although the involvement of D1R and D2R in dopamine-dependent learning has been suggested, the exact cellular bases for detecting transient dopamine signaling remain unclear. This review discusses recent cellular studies on the novel synaptic mechanisms for detecting dopamine transient signals associated with learning. Analyses of behavior based on these mechanisms have further revealed new behavioral aspects that are closely associated with these synaptic mechanisms. Thus, it is gradually possible to mechanistically explain behavioral learning via synaptic and cellular bases in rodents.

The plasticitome of cortical interneurons

Hebb postulated that, to store information in the brain, assemblies of excitatory neurons coding for a percept are bound together via associative long-term synaptic plasticity. In this view, it is unclear what role, if any, is carried out by inhibitory interneurons. Indeed, some have argued that inhibitory interneurons are not plastic. Yet numerous recent studies have demonstrated that, similar to excitatory neurons, inhibitory interneurons also undergo long-term plasticity. Here, we discuss the many diverse forms of long-term plasticity that are found at inputs to and outputs from several types of cortical inhibitory interneuron, including their plasticity of intrinsic excitability and their homeostatic plasticity. We explain key plasticity terminology, highlight key interneuron plasticity mechanisms, extract overarching principles and point out implications for healthy brain functionality as well as for neuropathology. We introduce the concept of the plasticitome - the synaptic plasticity counterpart to the genome or the connectome - as well as nomenclature and definitions for dealing with this rich diversity of plasticity. We argue that the great diversity of interneuron plasticity rules is best understood at the circuit level, for example as a way of elucidating how the credit-assignment problem is solved in deep biological neural networks.

neuroAIx-Framework: design of future neuroscience simulation systems exhibiting execution of the cortical microcircuit model 20× faster than biological real-time

Introduction

Research in the field of computational neuroscience relies on highly capable simulation platforms. With real-time capabilities surpassed for established models like the cortical microcircuit, it is time to conceive next-generation systems: neuroscience simulators providing significant acceleration, even for larger networks with natural density, biologically plausible multi-compartment models and the modeling of long-term and structural plasticity.

Methods

Stressing the need for agility to adapt to new concepts or findings in the domain of neuroscience, we have developed the neuroAIx-Framework consisting of an empirical modeling tool, a virtual prototype, and a cluster of FPGA boards. This framework is designed to support and accelerate the continuous development of such platforms driven by new insights in neuroscience.

Results

Based on design space explorations using this framework, we devised and realized an FPGA cluster consisting of 35 NetFPGA SUME boards.

Discussion

This system functions as an evaluation platform for our framework. At the same time, it resulted in a fully deterministic neuroscience simulation system surpassing the state of the art in both performance and energy efficiency. It is capable of simulating the microcircuit with 20× acceleration compared to biological real-time and achieves an energy efficiency of 48nJ per synaptic event.

Thalamocortical contribution to flexible learning in neural systems

Animal brains evolved to optimize behavior in dynamic environments, flexibly selecting actions that maximize future rewards in different contexts. A large body of experimental work indicates that such optimization changes the wiring of neural circuits, appropriately mapping environmental input onto behavioral outputs. A major unsolved scientific question is how optimal wiring adjustments, which must target the connections responsible for rewards, can be accomplished when the relation between sensory inputs, action taken, and environmental context with rewards is ambiguous. The credit assignment problem can be categorized into context-independent structural credit assignment and context-dependent continual learning. In this perspective, we survey prior approaches to these two problems and advance the notion that the brain's specialized neural architectures provide efficient solutions. Within this framework, the thalamus with its cortical and basal ganglia interactions serves as a systems-level solution to credit assignment. Specifically, we propose that thalamocortical interaction is the locus of meta-learning where the thalamus provides cortical control functions that parametrize the cortical activity association space. By selecting among these control functions, the basal ganglia hierarchically guide thalamocortical plasticity across two timescales to enable meta-learning. The faster timescale establishes contextual associations to enable behavioral flexibility, while the slower one enables generalization to new contexts.

Printed synaptic transistor-based electronic skin for robots to feel and learn

An electronic skin (e-skin) for the next generation of robots is expected to have biological skin-like multimodal sensing, signal encoding, and preprocessing. To this end, it is imperative to have high-quality, uniformly responding electronic devices distributed over large areas and capable of delivering synaptic behavior with long- and short-term memory. Here, we present an approach to realize synaptic transistors (12-by-14 array) using ZnO nanowires printed on flexible substrate with 100% yield and high uniformity. The presented devices show synaptic behavior under pulse stimuli, exhibiting excitatory (inhibitory) post-synaptic current, spiking rate-dependent plasticity, and short-term to long-term memory transition. The as-realized transistors demonstrate excellent bio-like synaptic behavior and show great potential for in-hardware learning. This is demonstrated through a prototype computational e-skin, comprising event-driven sensors, synaptic transistors, and spiking neurons that bestow biological skin-like haptic sensations to a robotic hand. With associative learning, the presented computational e-skin could gradually acquire a human body-like pain reflex. The learnt behavior could be strengthened through practice. Such a peripheral nervous system-like localized learning could substantially reduce the data latency and decrease the cognitive load on the robotic platform.

Striatum expresses region-specific plasticity consistent with distinct memory abilities

The striatum mediates two learning modalities: goal-directed behavior in dorsomedial (DMS) and habits in dorsolateral (DLS) striata. The synaptic bases of these learnings are still elusive. Indeed, while ample research has described DLS plasticity, little remains known about DMS plasticity and its involvement in procedural learning. Here, we find symmetric and asymmetric anti-Hebbian spike-timing-dependent plasticity (STDP) in DMS and DLS, respectively, with opposite plasticity dominance upon increasing corticostriatal activity. During motor-skill learning, plasticity is engaged in DMS and striatonigral DLS neurons only during early learning stages, whereas striatopallidal DLS neurons are mobilized only during late phases. With a mathematical modeling approach, we find that symmetric anti-Hebbian STDP favors memory flexibility, while asymmetric anti-Hebbian STDP favors memory maintenance, consistent with memory processes at play in procedural learning.

Canonical neural networks perform active inference

This work considers a class of canonical neural networks comprising rate coding models, wherein neural activity and plasticity minimise a common cost function-and plasticity is modulated with a certain delay. We show that such neural networks implicitly perform active inference and learning to minimise the risk associated with future outcomes. Mathematical analyses demonstrate that this biological optimisation can be cast as maximisation of model evidence, or equivalently minimisation of variational free energy, under the well-known form of a partially observed Markov decision process model. This equivalence indicates that the delayed modulation of Hebbian plasticity-accompanied with adaptation of firing thresholds-is a sufficient neuronal substrate to attain Bayes optimal inference and control. We corroborated this proposition using numerical analyses of maze tasks. This theory offers a universal characterisation of canonical neural networks in terms of Bayesian belief updating and provides insight into the neuronal mechanisms underlying planning and adaptive behavioural control.

Neural Circuits for Dynamics-Based Segmentation of Time Series

The brain must extract behaviorally relevant latent variables from the signals streamed by the sensory organs. Such latent variables are often encoded in the dynamics that generated the signal rather than in the specific realization of the waveform. Therefore, one problem faced by the brain is to segment time series based on underlying dynamics. We present two algorithms for performing this segmentation task that are biologically plausible, which we define as acting in a streaming setting and all learning rules being local. One algorithm is model based and can be derived from an optimization problem involving a mixture of autoregressive processes. This algorithm relies on feedback in the form of a prediction error and can also be used for forecasting future samples. In some brain regions, such as the retina, the feedback connections necessary to use the prediction error for learning are absent. For this case, we propose a second, model-free algorithm that uses a running estimate of the autocorrelation structure of the signal to perform the segmentation. We show that both algorithms do well when tasked with segmenting signals drawn from autoregressive models with piecewise-constant parameters. In particular, the segmentation accuracy is similar to that obtained from oracle-like methods in which the ground-truth parameters of the autoregressive models are known. We also test our methods on data sets generated by alternating snippets of voice recordings. We provide implementations of our algorithms at https://github.com/ttesileanu/bio-time-series.

Active inference leads to Bayesian neurophysiology

The neuronal substrates that implement the free-energy principle and ensuing active inference at the neuron and synapse level have not been fully elucidated. This Review considers possible neuronal substrates underlying the principle. First, the foundations of the free-energy principle are introduced, and then its ability to empirically explain various brain functions and psychological and biological phenomena in terms of Bayesian inference is described. Mathematically, the dynamics of neural activity and plasticity that minimise a cost function can be cast as performing Bayesian inference that minimises variational free energy. This equivalence licenses the adoption of the free-energy principle as a universal characterisation of neural networks. Further, the neural network structure itself represents a generative model under which an agent operates. A virtue of this perspective is that it enables the formal association of neural network properties with prior beliefs that regulate inference and learning. The possible neuronal substrates that implement prior beliefs and how to empirically examine the theory are discussed. This perspective renders brain activity explainable, leading to a deeper understanding of the neuronal mechanisms underlying basic psychology and psychiatric disorders in terms of an implicit generative model.

What Is the Role of Thalamostriatal Circuits in Learning Vocal Sequences?

Basal ganglia (BG) circuits integrate sensory and motor-related information from the cortex, thalamus, and midbrain to guide learning and production of motor sequences. Birdsong, like speech, is comprised of precisely sequenced vocal elements. Learning song sequences during development relies on Area X, a vocalization related region in the medial striatum of the songbird BG. Area X receives inputs from cortical-like pallial song circuits and midbrain dopaminergic circuits and sends projections to the thalamus. It has recently been shown that thalamic circuits also send substantial projections back to Area X. Here, we outline a gated-reinforcement learning model for how Area X may use signals conveyed by thalamostriatal inputs to direct song learning. Integrating conceptual advances from recent mammalian and songbird literature, we hypothesize that thalamostriatal pathways convey signals linked to song syllable onsets and offsets and influence striatal circuit plasticity via regulation of cholinergic interneurons (ChIs). We suggest that syllable sequence associated vocal-motor information from the thalamus drive precisely timed pauses in ChIs activity in Area X. When integrated with concurrent corticostriatal and dopaminergic input, this circuit helps regulate plasticity on medium spiny neurons (MSNs) and the learning of syllable sequences. We discuss new approaches that can be applied to test core ideas of this model and how associated insights may provide a framework for understanding the function of BG circuits in learning motor sequences.

SpikePropamine: Differentiable Plasticity in Spiking Neural Networks

The adaptive changes in synaptic efficacy that occur between spiking neurons have been demonstrated to play a critical role in learning for biological neural networks. Despite this source of inspiration, many learning focused applications using Spiking Neural Networks (SNNs) retain static synaptic connections, preventing additional learning after the initial training period. Here, we introduce a framework for simultaneously learning the underlying fixed-weights and the rules governing the dynamics of synaptic plasticity and neuromodulated synaptic plasticity in SNNs through gradient descent. We further demonstrate the capabilities of this framework on a series of challenging benchmarks, learning the parameters of several plasticity rules including BCM, Oja's, and their respective set of neuromodulatory variants. The experimental results display that SNNs augmented with differentiable plasticity are sufficient for solving a set of challenging temporal learning tasks that a traditional SNN fails to solve, even in the presence of significant noise. These networks are also shown to be capable of producing locomotion on a high-dimensional robotic learning task, where near-minimal degradation in performance is observed in the presence of novel conditions not seen during the initial training period.

Computational roles of intrinsic synaptic dynamics

Conventional theories assume that long-term information storage in the brain is implemented by modifying synaptic efficacy. Recent experimental findings challenge this view by demonstrating that dendritic spine sizes, or their corresponding synaptic weights, are highly volatile even in the absence of neural activity. Here, we review previous computational works on the roles of these intrinsic synaptic dynamics. We first present the possibility for neuronal networks to sustain stable performance in their presence, and we then hypothesize that intrinsic dynamics could be more than mere noise to withstand, but they may improve information processing in the brain.

Spine dynamics in the brain, mental disorders and artificial neural networks

In the brain, most synapses are formed on minute protrusions known as dendritic spines. Unlike their artificial intelligence counterparts, spines are not merely tuneable memory elements: they also embody algorithms that implement the brain's ability to learn from experience and cope with new challenges. Importantly, they exhibit structural dynamics that depend on activity, excitatory input and inhibitory input (synaptic plasticity or 'extrinsic' dynamics) and dynamics independent of activity ('intrinsic' dynamics), both of which are subject to neuromodulatory influences and reinforcers such as dopamine. Here we succinctly review extrinsic and intrinsic dynamics, compare these with parallels in machine learning where they exist, describe the importance of intrinsic dynamics for memory management and adaptation, and speculate on how disruption of extrinsic and intrinsic dynamics may give rise to mental disorders. Throughout, we also highlight algorithmic features of spine dynamics that may be relevant to future artificial intelligence developments.

Learning poly-synaptic paths with traveling waves

Traveling waves are commonly observed across the brain. While previous studies have suggested the role of traveling waves in learning, the mechanism remains unclear. We adopted a computational approach to investigate the effect of traveling waves on synaptic plasticity. Our results indicate that traveling waves facilitate the learning of poly-synaptic network paths when combined with a reward-dependent local synaptic plasticity rule. We also demonstrate that traveling waves expedite finding the shortest paths and learning nonlinear input/output mapping, such as exclusive or (XOR) function.

There are approximately 10¹¹ neurons with 10¹⁴ connections in the human brain. Information transmission among neurons in this large network is considered crucial for our behavior. To achieve this, multiple synaptic connections along a poly-synaptic network path must be adjusted coherently during learning. Because the previously proposed reward-dependent synaptic plasticity rule requires coactivation of presynaptic and postsynaptic neurons, learning can fail if a subset of neurons along a distant network path is inactive at the beginning of learning. We suggest that traveling waves that are initiated at an information source can mitigate this problem. We performed computer simulations of spiking neural networks with reward-dependent local synaptic plasticity rules and traveling waves. Our results show that this combination facilitates the learning and refinement of synaptic network paths. We argue that these features are a general biological strategy for maintaining and optimizing our brain function. Our research provides new insights into how complex neural networks in the brain form during learning and memory consolidation.

Reverse-Engineering Neural Networks to Characterize Their Cost Functions

This letter considers a class of biologically plausible cost functions for neural networks, where the same cost function is minimized by both neural activity and plasticity. We show that such cost functions can be cast as a variational bound on model evidence under an implicit generative model. Using generative models based on partially observed Markov decision processes (POMDP), we show that neural activity and plasticity perform Bayesian inference and learning, respectively, by maximizing model evidence. Using mathematical and numerical analyses, we establish the formal equivalence between neural network cost functions and variational free energy under some prior beliefs about latent states that generate inputs. These prior beliefs are determined by particular constants (e.g., thresholds) that define the cost function. This means that the Bayes optimal encoding of latent or hidden states is achieved when the network's implicit priors match the process that generates its inputs. This equivalence is potentially important because it suggests that any hyperparameter of a neural network can itself be optimized-by minimization with respect to variational free energy. Furthermore, it enables one to characterize a neural network formally, in terms of its prior beliefs.

Dopamine role in learning and action inference

This paper describes a framework for modelling dopamine function in the mammalian brain. It proposes that both learning and action planning involve processes minimizing prediction errors encoded by dopaminergic neurons. In this framework, dopaminergic neurons projecting to different parts of the striatum encode errors in predictions made by the corresponding systems within the basal ganglia. The dopaminergic neurons encode differences between rewards and expectations in the goal-directed system, and differences between the chosen and habitual actions in the habit system. These prediction errors trigger learning about rewards and habit formation, respectively. Additionally, dopaminergic neurons in the goal-directed system play a key role in action planning: They compute the difference between a desired reward and the reward expected from the current motor plan, and they facilitate action planning until this difference diminishes. Presented models account for dopaminergic responses during movements, effects of dopamine depletion on behaviour, and make several experimental predictions.

In the brain, chemicals such as dopamine allow nerve cells to ‘talk’ to each other and to relay information from and to the environment. Dopamine, in particular, is released when pleasant surprises are experienced: this helps the organism to learn about the consequences of certain actions. If a new flavour of ice-cream tastes better than expected, for example, the release of dopamine tells the brain that this flavour is worth choosing again.

However, dopamine has an additional role in controlling movement. When the cells that produce dopamine die, for instance in Parkinson’s disease, individuals may find it difficult to initiate deliberate movements. Here, Rafal Bogacz aimed to develop a comprehensive framework that could reconcile the two seemingly unrelated roles played by dopamine.

The new theory proposes that dopamine is released when an outcome differs from expectations, which helps the organism to adjust and minimise these differences. In the ice-cream example, the difference is between how good the treat is expected to taste, and how tasty it really is. By learning to select the same flavour repeatedly, the brain aligns expectation and the result of the choice. This ability would also apply when movements are planned. In this case, the brain compares the desired reward with the predicted results of the planned actions. For example, while planning to get a spoonful of ice-cream, the brain compares the pleasure expected from the movement that is currently planned, and the pleasure of eating a full spoon of the treat. If the two differ, for example because no movement has been planned yet, the brain releases dopamine to form a better version of the action plan. The theory was then tested using a computer simulation of nerve cells that release dopamine; this showed that the behaviour of the virtual cells closely matched that of their real-life counterparts.

This work offers a comprehensive description of the fundamental role of dopamine in the brain. The model now needs to be verified through experiments on living nerve cells; ultimately, it could help doctors and researchers to develop better treatments for conditions such as Parkinson’s disease or ADHD, which are linked to a lack of dopamine.

A two-compartment model of synaptic computation and plasticity

The synapse is typically viewed as a single compartment, which acts as a linear gain controller on incoming input. Traditional plasticity rules enable this gain control to be dynamically optimized by Hebbian activity. Whilst this view nicely captures postsynaptic function, it neglects the non-linear dynamics of presynaptic function. Here we present a two-compartment model of the synapse in which the presynaptic terminal first acts to filter presynaptic input before the postsynaptic terminal, acting as a gain controller, amplifies or depresses transmission. We argue that both compartments are equipped with distinct plasticity rules to enable them to optimally adapt synaptic transmission to the statistics of pre- and postsynaptic activity. Specifically, we focus on how presynaptic plasticity enables presynaptic filtering to be optimally tuned to only transmit information relevant for postsynaptic firing. We end by discussing the advantages of having a presynaptic filter and propose future work to explore presynaptic function and plasticity in vivo.

Supervised learning in spiking neural networks: A review of algorithms and evaluations

As a new brain-inspired computational model of the artificial neural network, a spiking neural network encodes and processes neural information through precisely timed spike trains. Spiking neural networks are composed of biologically plausible spiking neurons, which have become suitable tools for processing complex temporal or spatiotemporal information. However, because of their intricately discontinuous and implicit nonlinear mechanisms, the formulation of efficient supervised learning algorithms for spiking neural networks is difficult, and has become an important problem in this research field. This article presents a comprehensive review of supervised learning algorithms for spiking neural networks and evaluates them qualitatively and quantitatively. First, a comparison between spiking neural networks and traditional artificial neural networks is provided. The general framework and some related theories of supervised learning for spiking neural networks are then introduced. Furthermore, the state-of-the-art supervised learning algorithms in recent years are reviewed from the perspectives of applicability to spiking neural network architecture and the inherent mechanisms of supervised learning algorithms. A performance comparison of spike train learning of some representative algorithms is also made. In addition, we provide five qualitative performance evaluation criteria for supervised learning algorithms for spiking neural networks and further present a new taxonomy for supervised learning algorithms depending on these five performance evaluation criteria. Finally, some future research directions in this research field are outlined.

Single-cell transcriptomic evidence for dense intracortical neuropeptide networks

Seeking new insights into the homeostasis, modulation and plasticity of cortical synaptic networks, we have analyzed results from a single-cell RNA-seq study of 22,439 mouse neocortical neurons. Our analysis exposes transcriptomic evidence for dozens of molecularly distinct neuropeptidergic modulatory networks that directly interconnect all cortical neurons. This evidence begins with a discovery that transcripts of one or more neuropeptide precursor (NPP) and one or more neuropeptide-selective G-protein-coupled receptor (NP-GPCR) genes are highly abundant in all, or very nearly all, cortical neurons. Individual neurons express diverse subsets of NP signaling genes from palettes encoding 18 NPPs and 29 NP-GPCRs. These 47 genes comprise 37 cognate NPP/NP-GPCR pairs, implying the likelihood of local neuropeptide signaling. Here, we use neuron-type-specific patterns of NP gene expression to offer specific, testable predictions regarding 37 peptidergic neuromodulatory networks that may play prominent roles in cortical homeostasis and plasticity.

Bayesian Filtering with Multiple Internal Models: Toward a Theory of Social Intelligence

To exhibit social intelligence, animals have to recognize whom they are communicating with. One way to make this inference is to select among internal generative models of each conspecific who may be encountered. However, these models also have to be learned via some form of Bayesian belief updating. This induces an interesting problem: When receiving sensory input generated by a particular conspecific, how does an animal know which internal model to update? We consider a theoretical and neurobiologically plausible solution that enables inference and learning of the processes that generate sensory inputs (e.g., listening and understanding) and reproduction of those inputs (e.g., talking or singing), under multiple generative models. This is based on recent advances in theoretical neurobiology-namely, active inference and post hoc (online) Bayesian model selection. In brief, this scheme fits sensory inputs under each generative model. Model parameters are then updated in proportion to the probability that each model could have generated the input (i.e., model evidence). The proposed scheme is demonstrated using a series of (real zebra finch) birdsongs, where each song is generated by several different birds. The scheme is implemented using physiologically plausible models of birdsong production. We show that generalized Bayesian filtering, combined with model selection, leads to successful learning across generative models, each possessing different parameters. These results highlight the utility of having multiple internal models when making inferences in social environments with multiple sources of sensory information.

Robust random search with scale-free stochastic resetting

A new model of search based on stochastic resetting is introduced, wherein rate of resets depends explicitly on time elapsed since the beginning of the process. It is shown that rate inversely proportional to time leads to paradoxical diffusion which mixes self-similarity and linear growth of the mean-square displacement with nonlocality and non-Gaussian propagator. It is argued that such resetting protocol offers a general and efficient search-boosting method that does not need to be optimized with respect to the scale of the underlying search problem (e.g., distance to the goal) and is not very sensitive to other search parameters. Both subdiffusive and superdiffusive regimes of the mean-squared displacement scaling are demonstrated with more general rate functions.

Beyond the "Conceptual Nervous System": Can computational cognitive neuroscience transform learning theory?

In the last century, learning theory has been dominated by an approach assuming that associations between hypothetical representational nodes can support the acquisition of knowledge about the environment. The similarities between this approach and connectionism did not go unnoticed to learning theorists, with many of them explicitly adopting a neural network approach in the modeling of learning phenomena. Skinner famously criticized such use of hypothetical neural structures for the explanation of behavior (the "Conceptual Nervous System"), and one aspect of his criticism has proven to be correct: theory underdetermination is a pervasive problem in cognitive modeling in general, and in associationist and connectionist models in particular. That is, models implementing two very different cognitive processes often make the exact same behavioral predictions, meaning that important theoretical questions posed by contrasting the two models remain unanswered. We show through several examples that theory underdetermination is common in the learning theory literature, affecting the solvability of some of the most important theoretical problems that have been posed in the last decades. Computational cognitive neuroscience (CCN) offers a solution to this problem, by including neurobiological constraints in computational models of behavior and cognition. Rather than simply being inspired by neural computation, CCN models are built to reflect as much as possible about the actual neural structures thought to underlie a particular behavior. They go beyond the "Conceptual Nervous System" and offer a true integration of behavioral and neural levels of analysis.

Spatial Concept Learning: A Spiking Neural Network Implementation in Virtual and Physical Robots

This paper proposes an artificial spiking neural network (SNN) sustaining the cognitive abstract process of spatial concept learning, embedded in virtual and real robots. Based on an operant conditioning procedure, the robots learn the relationship of horizontal/vertical and left/right visual stimuli, regardless of their specific pattern composition or their location on the images. Tests with novel patterns and locations were successfully completed after the acquisition learning phase. Results show that the SNN can adapt its behavior in real time when the rewarding rule changes.

Multi-context blind source separation by error-gated Hebbian rule

Animals need to adjust their inferences according to the context they are in. This is required for the multi-context blind source separation (BSS) task, where an agent needs to infer hidden sources from their context-dependent mixtures. The agent is expected to invert this mixing process for all contexts. Here, we show that a neural network that implements the error-gated Hebbian rule (EGHR) with sufficiently redundant sensory inputs can successfully learn this task. After training, the network can perform the multi-context BSS without further updating synapses, by retaining memories of all experienced contexts. This demonstrates an attractive use of the EGHR for dimensionality reduction by extracting low-dimensional sources across contexts. Finally, if there is a common feature shared across contexts, the EGHR can extract it and generalize the task to even inexperienced contexts. The results highlight the utility of the EGHR as a model for perceptual adaptation in animals.

Theories of Error Back-Propagation in the Brain

This review article summarises recently proposed theories on how neural circuits in the brain could approximate the error back-propagation algorithm used by artificial neural networks. Computational models implementing these theories achieve learning as efficient as artificial neural networks, but they use simple synaptic plasticity rules based on activity of presynaptic and postsynaptic neurons. The models have similarities, such as including both feedforward and feedback connections, allowing information about error to propagate throughout the network. Furthermore, they incorporate experimental evidence on neural connectivity, responses, and plasticity. These models provide insights on how brain networks might be organised such that modification of synaptic weights on multiple levels of cortical hierarchy leads to improved performance on tasks.

Spiking Neurons Integrating Visual Stimuli Orientation and Direction Selectivity in a Robotic Context

Visual motion detection is essential for the survival of many species. The phenomenon includes several spatial properties, not fully understood at the level of a neural circuit. This paper proposes a computational model of a visual motion detector that integrates direction and orientation selectivity features. A recent experiment in the Drosophila model highlights that stimulus orientation influences the neural response of direction cells. However, this interaction and the significance at the behavioral level are currently unknown. As such, another objective of this article is to study the effect of merging these two visual processes when contextualized in a neuro-robotic model and an operant conditioning procedure. In this work, the learning task was solved using an artificial spiking neural network, acting as the brain controller for virtual and physical robots, showing a behavior modulation from the integration of both visual processes.

In vitro neural networks minimise variational free energy

In this work, we address the neuronal encoding problem from a Bayesian perspective. Specifically, we ask whether neuronal responses in an in vitro neuronal network are consistent with ideal Bayesian observer responses under the free energy principle. In brief, we stimulated an in vitro cortical cell culture with stimulus trains that had a known statistical structure. We then asked whether recorded neuronal responses were consistent with variational message passing based upon free energy minimisation (i.e., evidence maximisation). Effectively, this required us to solve two problems: first, we had to formulate the Bayes-optimal encoding of the causes or sources of sensory stimulation, and then show that these idealised responses could account for observed electrophysiological responses. We describe a simulation of an optimal neural network (i.e., the ideal Bayesian neural code) and then consider the mapping from idealised in silico responses to recorded in vitro responses. Our objective was to find evidence for functional specialisation and segregation in the in vitro neural network that reproduced in silico learning via free energy minimisation. Finally, we combined the in vitro and in silico results to characterise learning in terms of trajectories in a variational information plane of accuracy and complexity.

A Measure of Information Available for Inference

The mutual information between the state of a neural network and the state of the external world represents the amount of information stored in the neural network that is associated with the external world. In contrast, the surprise of the sensory input indicates the unpredictability of the current input. In other words, this is a measure of inference ability, and an upper bound of the surprise is known as the variational free energy. According to the free-energy principle (FEP), a neural network continuously minimizes the free energy to perceive the external world. For the survival of animals, inference ability is considered to be more important than simply memorized information. In this study, the free energy is shown to represent the gap between the amount of information stored in the neural network and that available for inference. This concept involves both the FEP and the infomax principle, and will be a useful measure for quantifying the amount of information available for inference.

Modulation of Spike-Timing Dependent Plasticity: Towards the Inclusion of a Third Factor in Computational Models

In spike-timing dependent plasticity (STDP) change in synaptic strength depends on the timing of pre- vs. postsynaptic spiking activity. Since STDP is in compliance with Hebb's postulate, it is considered one of the major mechanisms of memory storage and recall. STDP comprises a system of two coincidence detectors with N-methyl-D-aspartate receptor (NMDAR) activation often posited as one of the main components. Numerous studies have unveiled a third component of this coincidence detection system, namely neuromodulation and glia activity shaping STDP. Even though dopaminergic control of STDP has most often been reported, acetylcholine, noradrenaline, nitric oxide (NO), brain-derived neurotrophic factor (BDNF) or gamma-aminobutyric acid (GABA) also has been shown to effectively modulate STDP. Furthermore, it has been demonstrated that astrocytes, via the release or uptake of glutamate, gate STDP expression. At the most fundamental level, the timing properties of STDP are expected to depend on the spatiotemporal dynamics of the underlying signaling pathways. However in most cases, due to technical limitations experiments grant only indirect access to these pathways. Computational models carefully constrained by experiments, allow for a better qualitative understanding of the molecular basis of STDP and its regulation by neuromodulators. Recently, computational models of calcium dynamics and signaling pathway molecules have started to explore STDP emergence in ex and in vivo-like conditions. These models are expected to reproduce better at least part of the complex modulation of STDP as an emergent property of the underlying molecular pathways. Elucidation of the mechanisms underlying STDP modulation and its consequences on network dynamics is of critical importance and will allow better understanding of the major mechanisms of memory storage and recall both in health and disease.

SuperSpike: Supervised Learning in Multilayer Spiking Neural Networks

A vast majority of computation in the brain is performed by spiking neural networks. Despite the ubiquity of such spiking, we currently lack an understanding of how biological spiking neural circuits learn and compute in vivo, as well as how we can instantiate such capabilities in artificial spiking circuits in silico. Here we revisit the problem of supervised learning in temporally coding multilayer spiking neural networks. First, by using a surrogate gradient approach, we derive SuperSpike, a nonlinear voltage-based three-factor learning rule capable of training multilayer networks of deterministic integrate-and-fire neurons to perform nonlinear computations on spatiotemporal spike patterns. Second, inspired by recent results on feedback alignment, we compare the performance of our learning rule under different credit assignment strategies for propagating output errors to hidden units. Specifically, we test uniform, symmetric, and random feedback, finding that simpler tasks can be solved with any type of feedback, while more complex tasks require symmetric feedback. In summary, our results open the door to obtaining a better scientific understanding of learning and computation in spiking neural networks by advancing our ability to train them to solve nonlinear problems involving transformations between different spatiotemporal spike time patterns.

Error-Gated Hebbian Rule: A Local Learning Rule for Principal and Independent Component Analysis

We developed a biologically plausible unsupervised learning algorithm, error-gated Hebbian rule (EGHR)-β, that performs principal component analysis (PCA) and independent component analysis (ICA) in a single-layer feedforward neural network. If parameter β = 1, it can extract the subspace that major principal components span similarly to Oja’s subspace rule for PCA. If β = 0, it can separate independent sources similarly to Bell-Sejnowski’s ICA rule but without requiring the same number of input and output neurons. Unlike these engineering rules, the EGHR-β can be easily implemented in a biological or neuromorphic circuit because it only uses local information available at each synapse. We analytically and numerically demonstrate the reliability of the EGHR-β in extracting and separating major sources given high-dimensional input. By adjusting β, the EGHR-β can extract sources that are missed by the conventional engineering approach that first applies PCA and then ICA. Namely, the proposed rule can successfully extract hidden natural images even in the presence of dominant or non-Gaussian noise components. The results highlight the reliability and utility of the EGHR-β for large-scale parallel computation of PCA and ICA and its future implementation in a neuromorphic hardware.