Neural Networks

Equilibrium propagation (EP) is a biologically plausible alternative to backpropagation that has demonstrated competitive performance across a range of machine learning tasks. Recent work has extended EP to spiking neural networks (SNNs), leveraging leaky integrate-and-fire (LIF) neurons and spike-based plasticity rules to improve biological realism while maintaining strong performance. In this work, we propose an EP-based SNN framework that combines LIF neural dynamics with a predictive learning rule, replacing conventional spike-timing-dependent plasticity (STDP) with a learning rule more directly aligned with predictive coding principles. We evaluate the proposed model on multiple image classification benchmarks, including MNIST, KMNIST, and Fashion-MNIST, and compare its performance with a BP-trained LIF SNN baseline. Our results show that the proposed EP-based LIF model (EP+LIF) achieves competitive accuracy across datasets, with performance approaching that of the BP-trained counterpart (BP+LIF) while relying on a biologically motivated local learning rule. In addition, analysis of hidden-layer spiking activity reveals that EP+LIF produces more persistent hidden-state activity, whereas BP+LIF yields sparser spiking representations. These results demonstrate that predictive learning can support effective EP-based training in LIF spiking networks, while also highlighting differences in activity patterns that motivate future work on activity regulation and sparse spiking dynamics.

Improving long-term memory in artificial neural networks remains an open challenge. To address this, we developed a novel brain-inspired framework for memory prioritization based on the principle of emotional valence. Our framework includes: (i) a valence-weighted cross-entropy loss that scales the learning signal by the valence magnitude, analogous to neuromodulation; (ii) an amygdala-inspired module that learns high-dimensional valence embeddings; and (iii) a hippocampus-inspired module that integrates valence embeddings into the attention mechanism to modulate information retrieval. We demonstrated the generalization of our framework across spatial, episodic, and language-based memory tasks, consistently improving memory prioritization and long-term retention of high-salience information. In addition to improving long-term memory, we also showed that our framework can help mitigate the "lost-in-the-middle" problem in language modeling. More generally, this research provides further evidence of the potential of brain-inspired algorithms to advance the field of machine learning.

Categorisation is often treated as a form of compression: a high-dimensional stimulus space is reduced to a smaller set of behaviourally or cognitively useful classes. However, compression alone does not determine whether a category map is useful. The present manuscript develops an information-theoretic framework for evaluating categorisation in terms of both category complexity and target-relevant information preservation. Across a set of synthetic demonstrations, alternative category maps over the same stimulus space are shown to preserve different target variables, including identity, action, nuisance, and hierarchical category structure. The framework is then extended to learned visual representations by analysing layer-derived category maps from a pretrained ResNet-50 network applied to CIFAR-10 images. Two scenarios are compared: a clean-only object run and a pooled nuisance run containing clean, blurred, pixelated, and noise-perturbed images. The results show that category maps can have substantial entropy while preserving information about a variable that is not aligned with the specified target, and that the value of a categorisation depends on the target variable to be preserved. The manuscript argues that categorisation should therefore be evaluated not only by compression or separability, but by the information retained about a specified cognitive, behavioural, or computational target.

The learning dynamics of biological brains and artificial neural networks are of interest to both neuroscience and machine learning. A key difference between them is that neural networks are often trained from a randomly initialized state whereas each brain is the product of generations of evolutionary optimization, yielding innate structures that enable few-shot learning and inbuilt reflexes. Artificial neural networks, by contrast, require non-ethological quantities of training data to attain comparable performance. To investigate the effect of evolutionary optimization on the learning dynamics of neural networks, we combined algorithms simulating natural selection and online learning to produce a method for evolutionarily conditioning artificial neural networks, and applied it to both reinforcement and supervised learning contexts. We found the evolutionary conditioning algorithm, by itself, performs comparably to an unoptimized baseline. However, evolutionarily conditioned networks show signs of unique and latent learning dynamics, and can be rapidly fine-tuned to optimal performance. These results suggest evolution constitutes an inductive bias that tunes neural systems to enable rapid learning.

The awake brain is known to display asynchronous (AS) states during periods of attention and arousal, but the responsiveness properties of such states remain unclear. Here, we investigate this question using computational models of spiking networks of excitatory and inhibitory neurons, mimicking recurrently-connected networks in layer 2/3 of the cerebral cortex. The networks can generate a continuum of AS states, but with different responsiveness characteristics. By using a mean-field model to infer the dynamic properties of the system, we find that there are two families of AS states, which we call "underdamped" (UD) and "overdamped" (OD). Responsiveness is maximised at the transition between OD and UD states, which identifies a "working point" that may present advantageous computational properties.

In previous studies by the author on binocular vision with the asymmetric eye (AE), which models a healthy human eye with misaligned optical components, the results were primarily presented in the Rodrigues vector (RV) framework and supported by simulations and 3D visualizations in GeoGebras dynamic geometry environment. In this paper, the novel geometric kinematics of the human eye, that is, the eye with misaligned optics, and simplified assumptions about the eye rotations (the eyes translational movements are disregarded), are developed within the framework of rigid-body rotations. The originality of the analysis lies in a precise geometric decomposition of a full rotation of the eyes posture into a torsion-free rotation (the geodesic part) and a torsional rotation (the non-geodesic extension of the geodesic part). This decomposition is extended to the corresponding decomposition of the angular velocity. A novel derivation of the eyes angular velocity from the RV formulation of the eye kinematics is proposed.

Adolescence is an important developmental period during which there are diverse changes in the brain and behavior. Goal-directed behaviors and the component processes underlying those behaviors improve during adolescence, including working memory, response inhibition, and reinforcement learning. At the same time there is substantial pruning of excitatory connections in prefrontal cortex and ongoing myelination of axons. However, psychiatric disorders also become increasingly prevalent in late adolescence and early adulthood. In this study, we develop computational models that suggest a hypothesis for how the ongoing changes in the brain can give rise to the increased prevalence of psychiatric disorders. We show that both myelination and pruning during adolescence lead to attractor landscapes in which strongly encoded memories, driven by three-factor learning rules that modulate Hebbian plasticity, come to dominate the landscape of brain activity, at the expense of weakly encoded memories. Pruning and myelination lead to large, strong attractors which, if they are related to aversive emotions, can drive intrusive thoughts and compulsions in obsessive compulsive disorder, rumination in depression, and aversive memories in post-traumatic stress disorder. The link between pruning, myelination and the emergence of dominant attractors for emotionally salient memories is well supported by the models. The way these effects map onto forebrain circuits requires more work.

Cerebellar neural circuit dynamics rely on a rich repertoire of synaptic and excitability mechanisms, which are thought to determine network computation in physiological and pathological conditions. In this work, we develop and validate a biologically-grounded spiking neural network of the cerebellar cortex, embedding key mechanisms of cellular excitability and synaptic transmission, and assess their impact on signal processing. Neuronal input-output functions, short-term synaptic plasticity, receptor-specific kinetics, and NMDA channel voltage-dependent gating were calibrated against detailed multicompartmental models through automatic tuning procedures. Incorporating these realistic biological properties allowed the network model to simulate key features observed in recordings from acute cerebellar slices. The neuronal discharge and local field potentials elicited by mossy fiber stimulation faithfully reproduced the natural patterns with millisecond precision. Then, selective receptor switch-off revealed the contribution of NMDA, GABA, and AMPA receptors to the frequency-dependent input-output function of the granular layer and Purkinje cells, linking previous empirical findings to specific synaptic mechanisms. This model combines high computational performance with biological realism and offers a computationally efficient framework to investigate neurophysiological phenomena and the neural correlates of behavior in large-scale long-lasting simulations, such as those needed to address the neural underpinnings of learning and of cerebellar pathologies.

BackgroundThis study examines a competition-based model (C-model) designed to capture the temporal dynamics of successive brain microstates derived from electroencephalography (EEG) recordings during eyes-open conditions. The analyzed data were obtained from a public repository comprising microstate sequences from 60 sessions of a single subject [1]. When applied to microstate dynamics, the C-model posits a stochastic competition among neural circuits underlying the expression of individual microstates. MethodsThe model is formulated at a conceptual level (computational level in Marrs framework) and employs a geometric distribution to account for the long right tail of microstate duration distributions, interpreted as the probability of "failure" of the currently active microstate to persist. To account for the short-lived left tail, the model incorporates a transient increase in the stability of the currently active network, or equivalently, a temporary decrease in the activation probability of competing microstates (refractory period). ResultsThe model provides a good fit to the microstate duration distributions across all 60 sessions. One third of sessions showed microstate identity sequential dependency with respect to the previous microstates. DiscussionThese results suggest that the C-model captures key aspects of microstate temporal structure. Moreover, because microstate probabilities can be modulated by psychophysiological conditions--including the influence of previously active networks--the model may serve as a building block for more comprehensive neurobiological frameworks of neural and behavioral dynamics. In such frameworks, microstate sequences could emerge from structured competition and flow among neural networks supporting microstate expression.

The human visual system segments images using both high-level recognition mechanisms and low-level mechanisms that are largely independent of specific prior experience. The low-level mechanisms are essential for initiating recognition processes, and for learning to recognize new materials, objects, and contexts. Here we describe a hierarchical Bayesian observer (HBO) model of texture segmentation that is biologically plausible, takes into account the statistics of natural scenes, and does not depend on prior experience. The HBO model consists of five steps: local similarity grouping with local normalization, mutual similarity grouping (local grouping is strengthened if the neighboring regions are similar to the same set of other regions), transitive grouping (good continuation), confidence grouping (neighboring regions far from the same-different decision boundary guide grouping of regions near the decision boundary), and region grouping (similarity grouping of the regions from the initial segmentation). We find that a local similarity grouping process, trained to maximize accuracy, predicts human texture discrimination accuracy. We then find that the four additional steps accurately segment images with randomly shaped regions containing arbitrary natural textures. The success of the model depends on all the steps, but especially on local-similarity and transitive grouping. We also find that the transitive grouping allows correct segmentation of non-stationary texture regions (e.g., textures slanted in depth). Further, we find that when illumination varies across the image, local normalization enables both correct texture segmentation and estimation of illumination change. Finally, we find that unlike our model large state-of-the-art deep networks often fail on these stimuli.

Population analysis methods have become standard for navigating the complexity of neural data. However, these methods often assume a rate code, neglecting information encoded in the precise timing of spikes. Critically, additional information encoded in bursts of action potentials may be missed. Here, we develop a factor analysis method that disentangles the factors associated with bursts and individual spikes. This enables burst codes to be investigated directly from the structure of the data, without requiring external covariates. We demonstrate that analyzing firing rates alone obscures the latent structure and factors underlying bursts. Applying our method to simulated and experimental data, we show that it can infer the correct latent structure and be used to test for the presence of burst coding. By merging the population and burst coding perspectives, we provide a framework for linking changes in bursting to internal variables involved in attention, perception, and learning.

IntroductionAlthough neural activity is organized across temporal and spatial scales, the principles that determine the accuracy and fidelity of neural information representation across scales remain unclear. In particular, while recent empirical results have reported mesoscopic optimality in neural decoding, no theoretical accounts exist that explain when and why such intermediate scales emerge as optimal. Here, we develop an analytical framework to study the optimal temporal scale of information representation and its dependence on the dynamic structure of signal and noise in neural data. Materials and MethodsWe formulate a multiscale theoretical model in which neural population activity is represented by temporally encoded trial vectors at microscopic, mesoscopic, and macroscopic resolutions. Neural responses are modeled as class-dependent mean activations (signal) corrupted by temporally correlated noise, and decay rates of correlations in both signal and noise are varied parametrically. Representational quality at each scale is quantified using the sensitivity index (d-prime) for decoding condition from neural activity. ResultsWe derive closed-form expressions for the sensitivity index at each temporal scale. These expressions reveal the key roles of signal and noise correlations as the main determinants of condition decodability at all scales. Comparing expressions under various combinations of signal and noise correlations reveals two regimes. First, when signal and noise correlations are absent or persistent over time, the optimal resolution falls at one of two extremes: macroscale (resp. microscale) if signal correlations are stronger (resp. weaker) than noise correlations. In contrast, when both signal and noise correlations decay with temporal separation, temporal integration produces a nontrivial trade-off: moderate integration improves decodability by suppressing noise while preserving coherent signal, whereas excessive integration degrades signal and decodability. Therefore, only in the latter regime, mesoscopic representations emerge as optimal across a broad range of biologically plausible parameters. DiscussionThis work provides a theoretical explanation for how the optimal temporal scale of neural information representation depends on the interplay between signal and noise correlations and their temporal decay. Broadly, the framework establishes temporal integration as a principled mechanism linking multiscale neural dynamics to information representation and offers testable predictions across recording modalities and neural systems.

Understanding how neuronal populations interact to encode and transform sensory information is a fundamental challenge in computational neuroscience. Most existing studies, however, study neural encoding, behavioral readout, and functional connectivity as disjoint problems. Two-photon calcium imaging enables simultaneous recording of large neuronal ensembles in vivo, driven by diverse stimuli and eliciting distinct behaviors. However, extracting directional functional connectivity metrics as well as encoding and readout properties of neurons from such data remains difficult due to indirect and noisy observations of spiking activity, slow temporal dynamics, and the latent interplay between external stimuli and endogenous neural processes. Here, we introduce a unified conceptual and operational modeling and inference framework for directly extracting functional Granger causal (GC) effects between neurons, from external stimuli to neurons, and from neurons to behavior, from two-photon imaging data, in the sense of Granger. Inspired by the intersection information framework, we also identify neurons that encode features of sensory stimuli that inform behavioral readout. The resulting GC networks together with the taxonomy of functional sensori-behavioral relevance, which we call G-taxonomy, provides a powerful statistical analysis framework, enabled by the integration of several techniques including state-space modeling and inference, variational inference, and point processes. We applied the proposed framework to simulated and experimentally-recorded two-photon imaging from the mouse auditory cortex (A1) during both passive listening and active tone discrimination. Our simulation studies reveal significant improvement of our proposed methodology over existing techniques. Analysis of experimental data from the mouse A1 identifies distinct groups of cells with diverse sensori-behavioral relevance, as well as changes in functional connectivity associated with correct vs. incorrect behavior. In summary, this work provides a principled and data-driven methodology for uncovering directional interactions among the neurons, sensory stimuli, and behavior, all within the same statistical framework, offering new insights into how distributed cortical populations transform sensory inputs into behaviorally relevant representations. Author SummaryThe brain processes sensory inputs through the coordinated activity of large networks of neurons and produces readouts that elicit behavior. Understanding how information flows and is processed through these networks is a central goal of neuroscience. In this study, we present a new computational framework that identifies directional interactions among neurons in an ensemble as well as from sensory stimuli to neurons and from neurons to behavior. Utilizing the Granger formalism to identify directional effects, as opposed to common correlational measures, our framework extracts said effects directly from two-photon calcium imaging data. We tested our proposed method on both simulated data and recordings from the auditory cortex of mice during passive listening and active tone discrimination tasks. Our method revealed diverse groups of neurons in the auditory cortex with distinct functional roles and relevance to sensori-behavioral integration. Our framework provides a new way to study the flow of information in the brain and can be broadly applied to uncover neural computations across sensory and cognitive systems.

The phenomenon of splitting was originally observed in hamsters which, after prolonged exposure to constant light, exhibit two rest/wake cycles within a subjective day. Splitting is a consequence of the left and right suprachiasmatic nuclei (SCN) falling out of synchrony. While it is known that split activity is characterized by an antiphase relationship between the left and right SCN and between the core and shell within each hemisphere, the role of the commissural projections that connect the right and left SCN is not known. In the present study, we investigate the impact of the inter-hemispheric connections on the split and unsplit dynamics of a computational model of the bilateral SCN. Our model has 4 nodes corresponding to each right and left core and shell. We simulated our bilateral model under different lighting conditions and measured its period and the phase relationships among the 4 nodes. To further characterize the dynamics of the system, we performed a bifurcation analysis. We found that the bilateral model automatically splits unless entrained by bright light/dark cycles, or unless it has excitatory inter-hemispheric connections. This suggests that excitatory cross-connections may be important for freerunning behavior. We found that constant light of varying intensities transitions the model between split and unsplit activity only in very limited conditions, but the strength and polarity of the contralateral connections play a much greater role in this dynamical transition. These findings suggest that splitting may involve plasticity of the inter-hemispheric connections of the SCN.

Understanding how the brain represents hidden variables is a fundamental challenge. In navigation, the internal path-integration gain is often masked by external landmarks that override the path integrator. Path integration can recalibrate its gain when allothetic and idiothetic cues conflict, but the real-time dynamics of this process are hidden to direct observation. Here, we demonstrate that theta frequency provides an error signal between the observable hippocampal gain and the internal path-integration gain. Theta frequency decreased as conflict between landmark-driven hippocampal gain and path-integration gain increased and recovered as the path-integration gain recalibrated to the new gain. A continuous attractor model replicated these dynamics, suggesting that the theta-frequency drop is driven by the misalignment of allothetic and idiothetic inputs, reducing the excitatory drive to the network. Thus, theta frequency provides a real-time readout of internal gain-error signals, offering a novel methodology to estimate hidden cognitive variables through observable physiological oscillations.

Humans and animals often face risky situations that require decision-making. Such decisions can be high-risk, high-return at some times, and low-risk, low-return at other times, depending on the balance between optimism and pessimism. However, how this optimism-pessimism bias is regulated across contexts remains unclear. Here, we introduced a computational model of decision-making in a risk-taking task based on the free-energy principle, together with a machine-learning framework that inversely estimates cognitive updating and optimism-pessimism bias from behavioral data. Applying this framework to monkey behavioral data, we found that a monkey quickly and accurately recognized the degree of risk, while frequently switching between optimism and pessimism during the task. In addition, we identified a characteristic control rule for optimism-pessimism bias that is distinct from reward-dependent regulation. Our framework provided a principled tool for understanding the latent cognitive processes underlying risky decision-making in animals and humans.

Skin-tone labels are absent from public dermoscopy benchmarks such as the International Skin Imaging Collaboration (ISIC), making it impossible to audit whether clinical AI performs equitably across skin tones. While several recent works estimate skin tone automatically from clinical photography and selfies, we ask whether this approach is feasible on dermoscopy, the primary imaging modality of these benchmarks. To answer this, we make three main contributions. First, we release MST-Derm, a dual-rater Monk Skin Tone (MST) annotation benchmark on 500 ISIC 2018 images. Raters were given an explicit unrateable option for crops where the skin surrounding the lesion was too occluded to label confidently. We find that 60% of images were marked unrateable, yielding a 193-image consensus subset (quadratic-weighted Cohen's Kappa = 0.82). Second, we conduct a systematic feasibility study of three pixel-based MST annotation pipelines spanning the principal families in prior work: palette matching in perceptual colour space, robust colour statistics, and projection to a 1D colorimetric scalar. All three pipelines produce ordinal signal above chance (95% confidence intervals on quadratic-weighted Kappa exclude zero). However, ISIC 2018's extreme light-skin bias leaves 82% of the evaluation set at MST 2, giving a constant "always predict MST 2" baseline an accuracy floor the methods cannot overcome. To separate algorithmic signal from dataset bias, we evaluate on a class-balanced subset. The best method reaches quadratic-weighted Kappa = 0.43 against the trivial baseline of Kappa = 0.00, confirming the signal is genuine. Third, we diagnose this performance ceiling. We trace the bottleneck to two causes: dermoscopy's specialised illumination physically compresses the colour range on which lighter skin tones differ, and ISIC's dataset skew makes standard absolute-accuracy metrics uninformative. We conclude that while pixel-based colour features carry real MST signal on dermoscopy, current performance is insufficient for autonomous annotation. We release the benchmark, annotation protocol, all prediction runs, and analysis code to facilitate the development of robust skin-tone estimators, a vital prerequisite for accurately auditing fairness and mitigating bias in dermatological machine learning.

Human goal pursuit is often marked by persistent activity toward achieving an objective, as well as flexibility in switching objectives based on environmental demands. How humans balance the stability and flexibility necessary for goal pursuit is the key question of this study. We propose that goal pursuit generates dynamic attractor modes in policy landscapes that produce stability in goal pursuit. The attractor properties are modulated through progress monitoring, allowing for the flexibility necessary to switch objectives in favor of alternative goals. Through simulations and behavioral cloning of human participants performing an extended goal selection task, we show how dynamic modes can develop in the latent spaces of recurrent neural networks trained with reinforcement learning. We develop metrics to quantitatively assess the attractor qualities of dynamic modes, validating them against synthetically generated dynamical systems, and use them to investigate the context modulation of attractor modes during goal pursuit. We then proceed to develop a circuit-level account of goal persistence incorporating self-excitation and cross-inhibition as motifs for fast, self-sustaining dynamics modulated by slow, progress-integrating momentum and context signals. Lastly, we show that the switching costs experienced while managing multiple goals are an emergent property of resistance to the intrinsic dynamics of goal pursuit, thereby contributing a fresh perspective on the dynamics of extended goal pursuit.

Growing interest to describe the electrical behavior of glial cells, mainly astrocytes, in intact brain tissue poses more and more challenges to commonly accepted belief they only respond in a linear manner in uptake of the excess of extracellular potassium and maintenance of their network equipotentiality. Their highly conductive mutual interconnections via gap junction (GJ) connections introduce yet another class of nonlinear elements. As more studies report nonlinearities in membrane voltage Vm dependence of both, the membrane and junctional conductances, the need to formulate minimal dynamical models of their transient behavior is getting more acute. Since ODE models of coupled cells, even in simplest 1-d arrays, require simplified descriptions and small set of parameters, rare quantitative studies on glia makes the task even more difficult. This study attempts to qualify a self-coupled cell, or a glial cell coupled to fixed voltage as useful system for detecting the nature of instabilities and transitions coming from coupling. In a novel biophysical model of coupled astrocyte, we introduce nonlinear kinetics of deactivation for large junctional voltages for the first time. We found that N-shaped nonlinearities and corresponding fold structure in the vector field of isolated cell serves as a baseline on top of which coupling nonlinearities enrich the bifurcation picture. Numerical simulations of 1-d array of coupled astrocytes show that coupling increases the propensity of astrocytic Vm to bistability and front propagation. We believe that presented illustrations of possible effects of coupling nonlinearities will motivate neurobiologists to further explore their impact in disease. Significance statementTransient changes in membrane voltage of glial cells may produce significant transient voltage difference between directly coupled cells. Nonlinear steady-state conductance of their interconnection elements, the gap junctions, introduce nonlinear current profiles which are very difficult to measure and quantitate using the available methods due to marked permeability of the junctions and leakiness of glial membrane in general. We propose a minimal model of glial membrane extended with a self-coupled feedback loop, which under realistic simplifying assumptions could serve for qualitative analysis of the impact of coupling, on the stability of resting membrane voltage. Neuronal cells of the brain and spinal cord cannot exist and function without supportive and neuromodulatory functions of the diverse population of glial cells. This applies to virtually all physiological processes on cell level - from cell development, metabolic support, membrane signaling, slow molecular signal transduction, ion homeostasis, neurovascular coupling, myelination, to mention only a few, manifest neuro-glial interaction. Even though all glial cell types are interconnected, the most abundant ones, the astrocytes are massively interconnected by gap junctions to form ordered networks. Electrically, astrocytic networks display membrane voltage equipotentiality, which is considered system-wide resting state for given neuro-glial circuit or unit. With molecular and cellular substrates of glial connectivity being slowly elucidated, network science and dynamical modeling are slowly "invading" that area with many important issues left open. In this study using classical dynamical systems approaches we give indications how nonlinear intercellular coupling between astrocytes affects physiological resting state and its instabilities compared to isolated, uncoupled cell. We strongly believe the suggested minimal model could fill the gap in ODE modeling of neuro-glial circuits, within broadest scope of hypothesis-driven research in cell-level neuroscience.

Humans frequently make decisions in complex, high-dimensional environments, where identifying task-relevant information is critical for rapid behavior optimization. Humans outperform standard reinforcement learning agents in navigating such complexity, yet the cognitive strategies of humans remain unclear. To address this, we developed a novel multi-dimensional learning task in which only a subset of dimensions is reward-related. Crucially, unlike prior studies, subjects are uninformed of the true task dimensionality and have to identify them through exploration. This design closely mimics the ambiguity in real-world tasks. Our results have identified two stereotyped choice patterns that reveal "dimension-guided" strategies in exploration and exploitation. Cross-subject analyses suggest that dimension-guided exploration may promote the efficiency of reward-based learning. These findings indicate that humans leverage task dimensionality to guide exploration, and provide inspiration for improving exploration efficiency in AI agents.