The publication of Mainen and Sejnowski's 1995 seminal paper strongly renewed interest in how spike timing contributes to the neural code. In the 3 decades since then, considerable experimental and theoretical research has investigated the timescales at which spike timing contributes to the neural code. Here we review theoretical and experimental research of the last 30years aimed at defining conceptually and measuring operationally these timescales. By a critical review of the literature, we individuate six broad classes of timescales that have been conceptualized and operationalized: the maximal temporal precision of spiking that a neuron can achieve, the encoding time window (the time window containing the information-bearing spike times), the encoding timescale (the coarsest time resolution for measuring spikes without losing information), the maximal discrimination precision timescale (the smallest spike time difference that can be discriminated behaviorally), the encoding-readout intersection timescale (the maximal timing precision at which stimulus information encoded in neural activity is also actually read out to inform behavior), and the information consistency timescale (measuring the stability of information encoding over time). Together, this work has revealed short and long timescales that influence information coding and affect behavior. Short encoding timescales, from milliseconds to tens of milliseconds, are useful for sensory information encoding and perception. Long consistency timescales, ranging from hundreds of milliseconds to seconds, are useful for accumulating evidence and stabilizing decisions.

Since the advent of widely accessible AI tools, AI technology has been in high demand by businesses, academic researchers and individuals. Technology companies are building AI infrastructure at a rapid pace, and these facilities consume vast and growing resources, particularly electricity and water, with significant real and projected climate impacts. There is a need for new research initiatives to support long time horizon efforts to develop energy efficient computing capabilities to support the continued growth of AI infrastructure in a sustainable fashion. Such efficiency is required at both the hardware and software levels. Where can industry turn for examples of ultra-low power, energy efficient computing? We argue here that neurobiological principles offer rich and under-exploited sources of inspiration for energy efficient NeuroAI, and that new partnerships between industry and academia should be developed in this direction.

Targeted reaching movements and spatial coordination of footfall patterns are prime examples of spatial coordination of animal limbs. To explain this, both physiological and computational studies have suggested the use of movement primitives or the existence of an internal body representation. Since insects lack a dedicated posture-sensing organ or vestibular system (which vertebrates possess), it has been hypothesized that they derive high-level postural information from distributed low-level proprioceptive cues, integrated across their limbs. To test this possibility, we use a multi-layer spiking neural network to extract high-level information about limb movement and whole-body posture from information provided by distributed local proprioceptors. The preceding companion paper introduced the phasic-tonic encoding of joint angles by strictly local proprioceptive afferents, and high-accuracy encoding of joint angles and angular velocities in first-order interneurons. Here, we extend this model by second-order interneurons that encode movement primitives of single legs by coincidence detection from two or three leg-local inputs. By validation against annotated experimental data on whole-body kinematics of unrestrained stick insect locomotion, we show that modelled interneurons can signal particular step cycle phases, but also step cycle transitions such as leg lift-off. To indicate climbing behaviour, third-order interneurons encode body pitch relative to the substrate from position and motion of 6 times 3 local leg joints. Our results demonstrate that simple combinations of two or three position/velocity inputs from disjunct proprioceptor arrays are sufficient to encode high-order movement information about step cycle phases. The resulting movement primitive encoding may converge to represent particular locomotor states and whole-body posture.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00422-025-01029-x.

A fundamental question in neuroscience is how sensory signals are decoded from noisy cortical activity. We address this question in the olfactory system, decoding the route by which odorants arrive into the nasal cavity: through the nostrils (orthonasal), or through the back of the throat (retronasal). We recently showed with modeling and novel experiments on anesthetized rats that orthonasal versus retronasal modality information is encoded in the olfactory bulb (OB, a pre-cortical region). However, key questions remain: is modality information transmitted from OB to anterior piriform cortex (aPC)? How can this information be extracted from a much noisier cortical population with overall less firing? With simultaneous spike recordings of populations of neurons in OB and aPC, we show that an unsupervised and biologically plausible algorithm, Shared Covariance Decoding (SCD), can indeed linearly encode modality in low dimensional subspaces. Specifically, SCD improves encoding of ortho/retro in aPC compared to Fisher's linear discriminant analysis (LDA). Consistent with our theoretical analysis, when noise correlations between OB and aPC are low and OB well-encodes modality, modality in aPC tends to be encoded optimally with SCD. We observe that with several algorithms (LDA, SCD, optimal) the decoding accuracy distributions are invariant when GABA[Formula: see text] (ant-)agonists (bicuculline and muscimol) are applied to OB, which is consistent with invariance in population firing in aPC. Overall, we show modality information can be encoded efficiently in piriform cortex.

This study develops an event-based, energy-efficient control strategy for desynchronizing coupled neuronal networks using optimal control theory. Inspired by phase resetting techniques in Parkinson’s disease treatment, we incorporate stochasticity of the system’s dynamics into deterministic models to address neural system intrinsic noise. We use an advanced computational solver for nonlinear stochastic partial differential equations to solve the stochastic Hamilton–Jacobi–Bellman equation via level set methods for a single neuron model; this allows us to find control inputs which drive the dynamics close to the system’s phaseless set. When applied to coupled neuronal networks, these inputs achieve effective randomization of neuronal spike timing, leading to significant network desynchronization. Compared to its deterministic counterpart, our stochastic method can achieve considerable energy savings. The event-based control minimizes unnecessary charge transfer, potentially extending implanted stimulator battery life while maintaining robustness against variations in neuronal coupling strengths and network heterogeneities. These findings highlight the potential for developing energy-efficient neurostimulation techniques with implications for deep brain stimulation protocols. The presented computational framework could also be applied to other domains for which stochastic optimal control problems are prevalent.

Integrate-and-fire models are an important class of phenomenological neuronal models that are frequently used in computational studies of single neural activity, population activity, and recurrent neural networks. If these models are used to understand and interpret electrophysiological data, it is important to reliably estimate the values of the model’s parameters. However, there are no standard methods for the parameter estimation of Integrate-and-fire models. Here, we identify the model parameters of an adaptive integrate-and-fire neuron with temporally correlated noise by analyzing membrane potential and spike trains in response to a current step. Explicit formulas for the parameters are analytically derived by stationary and time-dependent ensemble averaging of the model dynamics. Specifically, we give mathematical expressions for the adaptation time constant, the adaptation strength, the membrane time constant, and the mean constant input current. These theoretical predictions are validated by numerical simulations for a broad range of system parameters. Importantly, we demonstrate that parameters can be extracted by using only a modest number of trials. This is particularly encouraging, as the number of trials in experimental settings is often limited. Hence, our formulas may be useful for the extraction of effective parameters from neurophysiological data obtained from standard current-step experiments.

Although the raison d’etre of the brain is the survival of the body, there are relatively few theoretical studies of closed-loop rhythmic motor control systems. In this paper we provide a unified framework, based on variational analysis, for investigating the dual goals of performance and robustness in powerstroke–recovery systems. To demonstrate our variational method, we augment two previously published closed-loop motor control models by equipping each model with a performance measure based on the rate of progress of the system relative to a spatially extended external substrate—such as a long strip of seaweed for a feeding task, or progress relative to the ground for a locomotor task. The sensitivity measure quantifies the ability of the system to maintain performance in response to external perturbations, such as an applied load. Motivated by a search for optimal design principles for feedback control achieving the complementary requirements of efficiency and robustness, we discuss the performance–sensitivity patterns of the systems featuring different sensory feedback architectures. In a paradigmatic half-center oscillator-motor system, we observe that the excitation–inhibition property of feedback mechanisms determines the sensitivity pattern while the activation–inactivation property determines the performance pattern. Moreover, we show that the nonlinearity of the sigmoid activation of feedback signals allows the existence of optimal combinations of performance and sensitivity. In a detailed hindlimb locomotor system, we find that a force-dependent feedback can simultaneously optimize both performance and robustness, while length-dependent feedback variations result in significant performance-versus-sensitivity tradeoffs. Thus, this work provides an analytical framework for studying feedback control of oscillations in nonlinear dynamical systems, leading to several insights that have the potential to inform the design of control or rehabilitation systems.

It is generally assumed that the brain uses something akin to sparse distributed representations. These representations, however, are high-dimensional and consequently they affect classification performance of traditional Machine Learning models due to the “curse of dimensionality”. In tasks for which there is a vast amount of labeled data, Deep Networks seem to solve this issue with many layers and a non-Hebbian backpropagation algorithm. The brain, however, seems to be able to solve the problem with few layers. In this work, we hypothesize that this happens by using Hebbian learning. Actually, the Hebbian-like learning rule of Restricted Boltzmann Machines learns the input patterns asymmetrically. It exclusively learns the correlation between non-zero values and ignores the zeros, which represent the vast majority of the input dimensionality. By ignoring the zeros the “curse of dimensionality” problem can be avoided. To test our hypothesis, we generated several sparse datasets and compared the performance of a Restricted Boltzmann Machine classifier with some Backprop-trained networks. The experiments using these codes confirm our initial intuition as the Restricted Boltzmann Machine shows a good generalization performance, while the Neural Networks trained with the backpropagation algorithm overfit the training data.

Neural circuits with multiple discrete attractor states could support a variety of cognitive tasks according to both empirical data and model simulations. We assess the conditions for such multistability in neural systems using a firing rate model framework, in which clusters of similarly responsive neurons are represented as single units, which interact with each other through independent random connections. We explore the range of conditions in which multistability arises via recurrent input from other units while individual units, typically with some degree of self-excitation, lack sufficient self-excitation to become bistable on their own. We find many cases of multistability-defined as the system possessing more than one stable fixed point-in which stable states arise via a network effect, allowing subsets of units to maintain each others' activity because their net input to each other when active is sufficiently positive. In terms of the strength of within-unit self-excitation and standard deviation of random cross-connections, the region of multistability depends on the response function of units. Indeed, multistability can arise with zero self-excitation, purely through zero-mean random cross-connections, if the response function rises supralinearly at low inputs from a value near zero at zero input. We simulate and analyze finite systems, showing that the probability of multistability can peak at intermediate system size, and connect with other literature analyzing similar systems in the infinite-size limit. We find regions of multistability with a bimodal distribution for the number of active units in a stable state. Finally, we find evidence for a log-normal distribution of sizes of attractor basins, which produces Zipf's Law when enumerating the proportion of trials within which random initial conditions lead to a particular stable state of the system.

Divisive normalization is a model of canonical computation of brain circuits. We demonstrate that two cascaded divisive normalization processors (DNPs), carrying out intensity/contrast gain control and elementary motion detection, respectively, can model the robust motion detection realized by the early visual system of the fruit fly. We first introduce a model of elementary motion detection and rewrite its underlying phase-based motion detection algorithm as a feedforward divisive normalization processor. We then cascade the DNP modeling the photoreceptor/amacrine cell layer with the motion detection DNP. We extensively evaluate the DNP for motion detection in dynamic environments where light intensity varies by orders of magnitude. The results are compared to other bio-inspired motion detectors as well as state-of-the-art optic flow algorithms under natural conditions. Our results demonstrate the potential of DNPs as canonical building blocks modeling the analog processing of early visual systems. The model highlights analog processing for accurately detecting visual motion, in both vertebrates and invertebrates. The results presented here shed new light on employing DNP-based algorithms in computer vision.