Sensory pathways differ widely across species and modalities in anatomy and function. Yet many solve variants of the same ancient, pervasive problem: quickly detecting meaningful changes in the world. Sometimes animals need to make these detections from noisy signals with millisecond precision to survive. We suggest how a single spiking neuron can address this selection pressure for quick detection. We present a simple relationship linking neural voltage dynamics to online hypothesis testing that we call the time constant rule. Assume that a neuron's membrane time constant equals the inverse of a change in its input rate. We show that this rule is then necessary and sufficient for its subthreshold membrane potential to represent a likelihood ratio of that rate change. Thresholding that likelihood ratio makes the neuron an online detector of input rate increases and, by extension, stimulus onset when rapid reaction is required. We tested the time constant rule by reanalyzing published recordings from eight second-order sensory neurons spanning olfaction, mechanoreception, electroreception, and audition, across both insects and vertebrates. In all cases, the measured membrane time constants were consistent with the predicted inverse relationship. These examples are intended as cross-system and cross-phyla consistency checks rather than a population-level validation study. When the time constant rule holds, it yields a direct and intuitive interpretation of neural computation. A neuron's voltage is an online likelihood ratio: its time constant sets the decay of statistical evidence, its threshold is a decision line, and each output spike is a time-stamped declaration that its input statistics have changed. The rule links membrane biophysics with the statistical computation and ecological function of a neuron. Its falsifiability invites direct experimental tests to determine whether it extends more broadly across sensory pathways and neuron classes.
Hyperdimensional computing (HDC) is gaining popularity as a lightweight alternative computing paradigm, where data are represented as long vectors (commonly of dimension 10,000) and processed using simple algebraic operations. In the two-part HDC classification pipeline, data are first encoded as a high-dimensional representation, then classified based on their similarity to a set of prototypes. Conventionally, the prototypes are trained while the encoder remains static. In this work, we propose a method of training a binary/bipolar HDC encoder, relying exclusively on native HDC integer and binary operations. Evaluations with our method show an average accuracy improvement of 2.13% over the untrained encoder on several classification data sets commonly used in HDC research, with no increase in model size or inference complexity.
There is a general consensus in theories of human speech recognition that humans engage in predictive processing during online speech processing. There are also claims that predictive processing is indicative of the operation of a predictive coding (PC) mechanism. PC is a generative, hierarchical feedback framework where feedback signals consist of input predictions, while feedforward signals consist primarily of prediction errors (PE). Some researchers have taken decreased neural signals when inputs conform to expectations as evidence for PC and claim that other possible explanatory frameworks (e.g., interactive activation) are incompatible with reported reductions in PE. However, these claims have been advanced using narrow-scope computational implementations of PC without known abilities to adequately (i.e., plausibly; simulate broader human spoken word recognition (SWR) behavioral phenomena. Here, we present the first known neurally and behaviorally adequate, mathematically formal neural network PC model of time-dependent human SWR. After demonstrating that the new model is able to emulate a fundamental, empirical signature of human SWR behavior (time course of lexical activation and competition, we compare model dynamics to neural performance targets that have been touted as hallmarks of PC in SWR. While the new model readily exhibits predictive processing (anticipatory activation of phonemes consistent with lexical knowledge) and reduced neural activity when inputs match expectations (a necessary component of PC more subtle patterns of neural activity under priming and noise conditions, previously proposed as diagnostic of PC do not emerge appreciably. This suggests that such patterns may not be hallmarks of PC. We discuss implications for PC-based theories of SWR.
A key challenge in visual object recognition is developing models that generalize from limited data while maintaining transparency in their decision making. We propose a biologically inspired model that addresses both issues by classifying images based on transformation-invariant local shape key features. Following the principles of the brain's what and where pathways, each feature is encoded by an image patch and its relative location in polar coordinates, enabling interpretable and robust comparisons between inputs and class prototypes. To mimic human concept learning, prototypes are selected using clustering, improving representativeness and generalization. Results show that our model achieves human-comparable performance, with an error rate between 1% and 2% on the MNIST data set when all training images are used as prototypes. In data-limited scenarios, where only a small number of prototypes are selected, our model consistently outperforms convolutional neural networks (CNNs). To evaluate out-of-distribution generalization, we use prototypes from MNIST and test both models on the ETL-1 data set, which differs in data distribution. Although CNN accuracy drops significantly under these conditions, our model maintains high accuracy, even with few prototypes, demonstrating strong robustness and greater capacity to generalize to unseen distributions, bringing it closer to human-like recognition capabilities.
Although evidence integration to the boundary model has successfully explained a wide range of behavioral and neural data in decision making under uncertainty, how animals learn and optimize the boundary remains unresolved. Here, we propose a model-free reinforcement learning algorithm for perceptual decisions under uncertainty that implements a sequential sampling process with an implicit decision boundary. Our model learns whether to commit to a decision given the available evidence or continue sampling information at a cost. We reproduced the canonical features of perceptual decision making such as dependence of accuracy and reaction time on evidence strength, modulation of speed-accuracy trade-off by payoff regime, and many others. By unifying learning and decision making within the same framework, this model can inspire a new look at the mechanisms of flexibility in context-dependent changes of behavior.
This letter introduces a novel method for discrete-sequence classification within the hyperdimensional computing (HDC) paradigm. The method, inspired by the concept of hidden Markov models (HMM), implements mechanisms that effectively address the major challenges arising in analyzing sequences generated by real-world processes, such as variable length, misalignment, symbol repetitions, omissions, or insertions. All algebraic transformations of input data in HMM-based sequence classification are replaced by bitwise operations on appropriately derived hyperdimensional binary vectors (hypervectors). The proposed hypervector transformation pipeline mirrors the algebraic manipulations of the HMM paradigm and involves a procedure that prevents information decay when processing longer sequences. Since only binary bit-wise operations on hypervectors are used in computations, the proposed method is well suited for hardware implementation, especially in the form of custom in-memory computing VLSI devices. The experimental evaluation of the proposed method, conducted on both artificial and real-world data sets, demonstrates its superiority over existing HDC-based sequence analysis approaches, yielding noticeable gains in classification accuracy. We also demonstrate that the proposed approach, like most HDC-based sequence classification methods, is robust against adverse physical, internal, or external factors that induce bit flips in the underlying hardware, which could entirely corrupt the outcomes of classical sequence-analysis methods, such as HMMs.
Identifying phase coupling from electrophysiological signals recorded by multiple electrodes, such as electroencephalogram (EEG) and electrocorticography (ECoG), helps neuroscientists and clinicians understand underlying brain structures or mechanisms. From a statistical perspective, these signals are multidimensional circular measurements that are correlated with one another and can be effectively modeled using a torus graph model designed for circular random variables. Using the torus graph model avoids the issue of detecting spurious correlations. However, the naive estimation of this model tends to lead to a dense network structure that is difficult to interpret. Therefore, to enhance the interpretability of the brain network structure, this review proposes a sparse estimation method for the torus graph model using regularized score matching combined with information criteria. In numerical simulations, our method successfully recovered the true dependence structure from a synthetic data set. Furthermore, we present analyses of two real data sets, one involving human EEG and the other marmoset ECoG, demonstrating that our method can be widely applied to phase-coupling analysis across different types of neural data. We found that when we used our proposed method, the modularity of the estimated network structure revealed more resolved brain structures and demonstrated differences in trends among individuals.
In many domain adaptation tasks, the source and target domains share an identical feature space, so the domain gap arises only from the distributional shift. In practice; however, new-target-only features (e.g., sensors added after training) often become available at test time, violating the shared feature space assumption and invalidating most existing methods. We address this setting with Label-Aware and Graph-Based Fused Gromov-Wasserstein Optimal Transport (LAGB-FGW), focusing on a transductive domain adaptation scenario, in which the entire unlabeled target data set is available during training, and predictions are jointly inferred for all target samples. LAGB-FGW (1) embeds label discrepancy directly into the source metric, (2) constructs a K-NN graph on the full target feature space to capture structure introduced by the additional features, and (3) jointly solves standard Optimal Transport (OT) and Gromov-Wasserstein OT, thereby transferring labels using both the common and the additional features. We validate LAGB-FGW on four synthetic benchmarks and the HAR70+ human-activity data set, and LAGB-FGW consistently outperforms all baselines, highlighting the advantage of combining source label information with graph-based structural cues when additional target features are available only at test time.
Active inference, a neurally inspired model for inferring actions based on the free energy principle (FEP), has been proposed as a unifying framework for understanding perception, action, and learning in the brain. Active inference has previously been used to model ecologically important tasks such as navigation and planning, but scaling it to solve complex large-scale problems in real-world environments has remained a challenge. Inspired by the existence of multiscale hierarchical representations in the brain, we propose a model for planning actions based on hierarchical active inference. Our approach combines a hierarchical model of the environment with successor representations for efficient planning. We present results demonstrating (1) how lower-level successor representations can be used to learn higher-level abstract states, (2) how planning based on active inference at the lower level can be used to bootstrap and learn higher-level abstract actions, and (3) how these learned higher-level abstract states and actions can facilitate efficient planning. We illustrate the performance of the approach on several planning and reinforcement learning (RL) problems, including a variant of the well-known four rooms task, a key-based navigation task, a partially observable planning problem, the mountain car problem, and PointMaze, a family of navigation tasks with continuous state and action spaces. Our results represent, to our knowledge, the first application of learned hierarchical state and action abstractions to active inference in FEP-based theories of brain function.
In this letter, we propose Bayesian Tucker decomposition (BTuD) in which the residuals are supposed to follow gaussian distribution analogous to linear-regression. Although we have proposed an algorithm to perform the proposed BTuD, the conventional higher-order orthogonal iteration can generate Tucker decomposition consistent with the present implementation. Using the proposed BTuD, we can perform unsupervised feature selection successfully applied to various synthetic data sets, global coupled maps with randomized coupling strength, and gene expression profiles. Thus, we can conclude that our newly proposed unsupervised feature selection method is promising. In addition to this, BTuD-based unsupervised feature extraction (FE) is expected to coincide with TD-based unsupervised FE that were previously proposed and successfully applied to a wide range of problems.
This letter introduces a novel algorithm for training deep neural networks with many nonlinear layers. Our method uses an approximated integrated gradient that is averaged over the range of the weight update to more accurately capture the loss change resulting from parameter updates. Unlike standard gradients, this average gradient improves learning efficiency in certain scenarios. We incorporate the approximated average gradient into RMSProp and compare the resulting algorithm to conventional RMSProp and Adam. We evaluate the approach on deep models lacking skip connections, such as those with many nonlinear activations and no residual structure, where traditional methods typically encounter difficulties. These models that focus on extracting high-order features create a loss landscape more akin to that of a biological brain. Our method requires significantly fewer iterations to reach a target training loss on MNIST, Fashion MNIST, and IMDb benchmarks for both convolutional and fully connected architectures across different initialization schemes. While our approach incurs moderately higher computational and memory costs compared to standard RMSProp, its performance on shallow models remains comparable. Nevertheless, our main contributions are (1) introducing the average gradient concept as an efficient alternative to computing high-order derivatives, (2) offering a novel factorization formula for approximating the average gradient, accompanied by a formal derivation, and (3) showing an example algorithm that leverages this formula to enhance the efficiency of RMSProp for some models, as validated by our evaluation.
Movement duration, a fundamental aspect of motor control, is often viewed as a preprogrammed parameter requiring dedicated selection mechanisms. An alternative view posits that movement duration emerges from the control policy itself. Here, we demonstrate, using infinite horizon optimal feedback control (IHOFC) and nonlinear limb dynamics, that this alternative hypothesis successfully captures diverse aspects of human reaching behavior, including trade-offs between movement duration and task parameters. Specifically, we reproduced the modulation of movement duration with varying reach distances and accuracy (Fitts's law) in the presence of nonlinear dynamics, and extended the infinite horizon framework to include the effect of rewards and biomechanical costs. Furthermore, our model also featured a temporal evolution of feedback responses to perturbations that resembles experimental observations and naturally accounted for motor decisions observed when participants select one among multiple goals in dynamic environments. Together, these developments show that in many cases, movement duration may not need to be specified a priori, but instead could result from task-dependent control policies. This framework validates a candidate explanation for varied movement durations, which invites a reconsideration of the nature and strength of evidence for the finite horizon formulation.
In human motor coordination, learning to coactivate multiple muscles at once to achieve distinct target combinations of forces or tasks remains a fundamental area of study. Task interference, where training on one task degrades performance on previously learned tasks, can slow motor learning. However, the neural mechanisms that reduce interference are not fully understood. We hypothesized that the structure of planning noise, specifically its signal-dependent nature, significantly shapes learning dynamics and limits interference within motor learning systems that rely on variability for exploration. To test this hypothesis, we developed a three-layer neural network model of muscle coordination informed by key neuroanatomical and neurophysiological principles and simulated learning for producing various combinations of muscle forces. Synaptic weights were stochastically altered from trial to trial with either fixed-variance planning noise (FVPN), where each connection's variance was fixed during learning, or signal-dependent planning noise (SDPN), where noise variance depended on the neuron population activity. Weights were reinforced when they reduced output error relative to target forces. An execution noise term, applied to the motor output, modeled peripheral motor variability. However, the learning rule was not informed about how much of the output corresponded to peripheral or central variability. Our results showed that SDPN improved both the rate and accuracy of multitask learning by reducing task interference compared to FVPN across network sizes, training schedules, and execution noise levels. SDPN achieved this by concentrating neural plasticity within the neuron populations engaged by the current task rather than modifying the entire network. This signal-dependent plasticity allowed multiple motor primitives to form, stabilize, and be reused for new tasks. The model replicated the well-documented benefit of interleaved versus blocked training in motor learning. As a computational proof of concept, this work suggests that SDPN can benefit multitask motor training by facilitating the formation of motor primitives.
This article introduces a family of multiclass linear perceptron classifiers with a multiplicative margin mechanism (MMPerc), as an alternative to standard margin-free and additive margin perceptrons. The multiplicative formulation enforces classification confidence by requiring the true class score to exceed that of competing classes by a specified fraction of itself rather than by a fixed additive threshold. This avoids dependence on score magnitudes arising from varied norms of data and class weight vectors. We propose several architectural and algorithmic variants of MMPerc, derive associated loss functions and mistake bounds for both linearly separable and nonseparable data, and analyze key design considerations, including bias, margin threshold selection, and training modes. Extensive experiments on synthetic and real data sets show that MMPerc classifiers typically outperform the standard perceptron, as well as classic baselines such as support vector machines and ridge classifiers. Owing to their simplicity, minimalistic design, and computational efficiency, MMPerc classifiers are promising candidates for conventional machine learning tasks, linear evaluation of deep neural networks, integration with hyperdimensional computing and vector symbolic architecture representations, and deployment in resource-constrained applications.
This article explores how simple reinforcement learning algorithms might be implemented by the anatomy of the cerebellum. In doing this, we highlight which anatomical and physiological details are most important for assessing algorithmic fit, and we discuss which algorithm components are easiest to accommodate in a neural system. We describe hypothetical cerebellar implementations of four reinforcement learning algorithms and discuss the anatomical plausibility of the various components required. We show how one of the algorithms can learn to generate short sequences of actions without continuous information on the resulting changes to the environment. We finish with simulations that illustrate the way that the algorithms learn to solve the problem of balancing an inverted pendulum, commonly known as the cart-pole problem. We highlight two physiological features: reward signals and combining information across time, that indicate that some sort of reinforcement learning adaptation may be taking place. We also describe why the commonly used algorithmic feature, an eligibility trace, presents particular problems to implement in known neural anatomy.
Force learning is a learning method for generating various types of complex dynamics in recurrent neural networks (RNNs), which is related to the reservoir computing (RC). RC uses an RNN called reservoir whose synaptic weights are randomly generated and fixed during learning. Force learning trains these synaptic weights inside the reservoir networks. Although force learning can be used as an effective tool for machine learning, possibilities of its realization in the brain are not often discussed. Here, in order to consider the possibilities of its realization in the brain, force learning is applied to an excitatory and inhibitory (E-I) network that models the cerebral cortex. A multimodule network composed of excitatory and inhibitory neurons is defined, and a readout is put outside, similar to a conventional reservoir. The output of this network is calculated at the readout as a linear combination of the filtered average firing rates of the excitatory neurons in the modules. Feedback connections that provide output back to the excitatory neurons in the modules with random strength are also added to this network. This network typically shows transitive chaotic synchronization, in which synchronizing modules are rearranged chaotically and intermittently. Under such conditions, our E-I network is trained to generate sinusoidal periodic signals for simplicity with force learning. When adjusting the E-I activity, it is observed that the efficiency of force learning is maximized at an optimal E-I balance near an edge of chaos. These results imply that the cooperation of excitatory and inhibitory neurons is required when force learning works effectively in the brain, although usual reservoir networks don't distinguish these two kinds of neurons.
Neuromodulators are critical controllers of neural states, with dysfunctions linked to various neuropsychiatric disorders. Although many biological aspects of neuromodulation have been studied, the computational principles underlying how neuromodulation of distributed neural populations controls brain states remain unclear. In contrast to external contextual inputs, neuromodulation can act as a single scalar signal that is broadcast to a vast population of neurons. We model the modulation of synaptic weight in a recurrent neural network model and show that neuromodulators can dramatically alter the function of a network, even when highly simplified. We find that under structural constraints like those in brains, this provides a fundamental mechanism that can increase the computational capability and flexibility of a neural network. Diffuse synaptic weight modulation enables storage of multiple memories using a common set of synapses that are able to generate diverse, even diametrically opposed, behaviors. Our findings help explain how neuromodulators unlock specific behaviors by creating task-specific hyperchannels in neural activity space and motivate more flexible, compact and capable machine learning architectures.
Exploring the dynamics and complexity of brain signal is critical to advancing our understanding of brain function. Recent fMRI studies have revealed links between BOLD signal variability or complexity with static/dynamics features of functional brain networks (FBN). However, the association between variability/complexity and regional centrality is still understudied. Here we investigate the association between variability/complexity and static/dynamic nodal features of FBN using graph theory analysis with fMRI BOLD data acquired during naturalistic movie watching. We found that variability positively correlated with fine-scale complexity but negatively correlated with coarse-scale complexity. Specifically, regions with high centrality and clustering coefficient were related to less variable but more complex signal. Similar relationships persisted for dynamic FBN, but the associations with certain aspects (e.g., eigenvector centrality) of regional centrality dynamics became insignificant. Our findings demonstrate that the relationship between BOLD signal variability and static/dynamic FBN with BOLD signal complexity depends on the temporal scale of signal complexity and that time-varying features of FBN reflect the complexities of how BOLD signal variability/complexity coevolve with dynamic FBN.
Presynaptic axon terminals maintain in their cytosol an almost constant level of adenosine triphosphate (ATP) to safeguard neurotransmission during varying workloads. In the study reported in this letter, it is argued that the vesicular release of neurotransmitter and the recycling of transmitter via astrocytes may itself be a mechanism of ATP homeostasis. In a minimal metabolic model of a presynaptic axon bouton, the accumulation of glutamate into vesicles and the activity-dependent supply of its precursor glutamine by astrocytes generated a steady-state level of ATP that was independent of the workload. When the workload increased, an enhanced supply of glutamine raised the rate of ATP production through the conversion of glutamate to the Krebs cycle intermediate α-ketoglutarate. The accumulation and release of glutamate, on the other hand, acted as a leak that diminished ATP production when the workload decreased. The fraction of ATP that the axon spent on the release and recycling of glutamate was small (4.7%), irrespective of the workload. Increasing this fraction enhanced the speed of ATP homeostasis and reduced the futile production of ATP. The model can be extended to axons releasing other, or coreleasing multiple, transmitters. Hence, the activity-dependent formation and release of neurotransmitter may be a universal mechanism of ATP homeostasis.
Biological neural networks self-organize according to local synaptic modifications to produce stable computations. How modifications at the synaptic level give rise to such computations at the network level remains an open question. Pehlevan et al. (2015) proposed a model of a self-organizing neural network with Hebbian and anti-Hebbian synaptic updates that implements an algorithm for principal subspace analysis; however, global stability of the nonlinear synaptic dynamics has not been established. Here, for the case that the feedforward and recurrent weights evolve at the same timescale, we prove global stability of the continuum limit of the synaptic dynamics and show that the dynamics evolve in two phases. In the first phase, the synaptic weights converge to an invariant manifold where the neural filters are orthonormal. In the second phase, the synaptic dynamics follow the gradient flow of a nonconvex potential function whose minima correspond to neural filters that span the principal subspace of the input data.
