Introduction Generative mechanisms of perception such as predictive coding are used to explain how the brain perceives the world; such mechanisms are often experimentally probed using “deviant” stimuli that violate established patterns (including mismatch negativity), which also elicit responses related to lower-level processes such as stimulus-specific adaptation. However, little is still known about brain responses that indicate the strength of sensory predictions or reinforcement of sensory representations. Repetition positivity (RP) is a positive polarity evoked potential that gradually increases with each repetition of a stimulus, and is thought to reflect progressive strengthening of auditory sensory memory and/or habituation to repetitive stimuli. The aim of this study was to compare RP that follows a change in stimulus frequency with that following a change in stimulus intensity, the latter having not previously been studied. Methods We used roving sequences of isochronous 5 kHz pure tones (300 ms duration, 300ms inter-stimulus interval), which changed in frequency by 1 kHz (Experiment 1) or in intensity by 12 dB (Experiment 2) after every 30 stimuli. All changes were roving, such that an increase would be followed by a decrease, and vice versa. Results Event-related potentials recorded with EEG indicated that frequency changes in either direction were followed by RP, whilst only intensity increases were followed by RP, and only a weak visual trend toward RP was apparent for intensity decreases. Observed RP was best explained by a logarithmic function over successive stimuli. Conclusions RP robustly follows increases, but not necessarily decreases, in stimulus intensity, which appears smaller in amplitude than that elicited by similarly salient frequency changes, and reaches a plateau sooner. These observations offer insight into how intensity is processed similarly yet differently to other sensory attributes in an adaptive or predictive coding framework, and might have future utility in the study of clinical conditions related to aberrant predictive mechanisms.

Cardio-respiratory interactions in interoceptive perception: The role of heartbeat-modulated cortical oscillations.

Functional neurological disorder (FND) is a disabling neuropsychiatric condition characterized by altered voluntary motor or sensory functions and cognitive symptoms. Unlike other neurological disorders, FND is not caused by structural brain damage but by disruptions across brain networks involved in agency, attention, emotion processing, sensory-motor control, and interoception. These alterations align with alterations in predictive coding, which propose that abnormal prior beliefs override sensory input, contributing to symptom generation. Early-life trauma is a significant risk factor, interacting with genetic and epigenetic vulnerabilities that influence emotional regulation, stress sensitivity, and brain connectivity. Psychiatric comorbidities are also common and may affect symptom severity and prognosis. This review synthesizes recent research to clarify the complex mechanisms underlying FND by integrating neurobiological, environmental, and psychological factors. By doing so, we aim to advance the understanding of FND pathophysiology and promote a more comprehensive conceptual framework that highlights the role of individual vulnerability.

Understanding semantic impairments in schizophrenia from a predictive coding perspective.

The pulvinar nucleus and its role in cognitive functions.

Brain‐computer interface (BCI) plays an important role in various fields, such as neuroscience, rehabilitation, and machine learning. The silent BCI, which can reconstruct inner speech from neural activity, holds great promise for aphasia patients. In this paper, we design an imagined Chinese speech experimental paradigm based on initials and finals and collect raw signals from eight healthy participants by using 64‐channel scalp electroencephalograms. Linear predictive coding (LPC) and mel frequency cepstral coefficients (MFCC), which are classical algorithms in the field of speech recognition, are used to extract distinguishing features for speech classification and reconstruction. Besides, the phase‐lock value (PLV) is introduced to enrich the feature information. We choose support vector machine (SVM), linear discriminant analysis (LDA), decision tree (DT), and LogitBoost (LB) for binary classification in several different cases. Two‐channel selection (CS) based on Broca’s area and Wernicke’s area of the brain is also introduced in the paper. The highest imaginary speech decoding accuracy reaches 84.38%, which demonstrates the effectiveness of the feature engineering. In addition, the comparative analysis is conducted with deep learning methods specifically designed for small sample scenarios. This study offers a novel systematic approach for the research of Chinese speech imagination BCI.

Continual learning with a predictive coding based classifier

Integrated estimation and control present an ongoing challenge for robotic systems. Because controllers depend on data derived from measured states and parameters, which are often subject to uncertainties and noise. The suitability of frameworks depends on the complexity of the task and the constraints of computational resources. They must strike a balance between computational efficiency for rapid responses while maintaining accuracy and robustness for safe and reliable missions. This study capitalizes on recent advancements in neuromorphic computing tools, especially spiking neural networks (SNNs), and their applications in robotic and dynamical systems. We present a learning-free framework featuring a recurrent network of leaky integrate-and-fire (LIF) neurons, designed to mimic a linear quadratic regulator (LQR) provided by a robust filtering strategy called extended modified sliding innovation filter (EMSIF). Thus, our proposed framework benefits from the robustness of EMSIF and the computational efficiency of SNN. The weight matrices of SNN are tailored to match the desired system model, eliminating the need for training. Moreover, the network leverages a biologically plausible firing rule akin to predictive coding. Furthermore, in the presence of various uncertainties, the SNN-LQR-EMSIF compared with non-spiking LQR-EMSIF, and the optimal strategy called linear quadratic Gaussian (LQG) based on extended Kalman filter. We evaluate their performance in a workbench problem and, next in the satellite rendezvous maneuver implement the Clohessy-Wiltshire (CW) model. Results demonstrated that the SNN-LQR-EMSIF achieves acceptable performance in terms of computational efficiency, robustness, and accuracy, positioning it as a promising approach for addressing the challenges of Integrated estimation and control in dynamic systems.

Background. Physical limitations of a person sometimes make it impossible to operate a computer with only a keyboard and mouse, Aims. One tool that can be used is a voice command, which is part of speech recognition technology. Methods. The voice signal will be normalized first, and then the coefficient values will be calculated using the Linear Predictive Coding (LPC) and Fast Fourier Transform (FFT) methods. After the coefficient value is obtained, recognition is performed using the backpropagation method of the Artificial Neural Network. Conclusion. The artificial neural network backpropagation method is used because it can adjust its own weights and produce error values that we can determine, thereby improving accuracy. Implementation. This study implements a voice command system using MARF as its speech engine and Java as its programming language.

BackgroundVoice biomarkers can effectively indicate early Mild Cognitive Impairment (MCI)1‐10. This study evaluates how each speech type performs in screening through active tasks. We also examine which combinations of these tasks best detect MCI.MethodWe designed three tasks to that end: Scripted Reading (Task 1) and Picture‐Based Question and Answer (Task 2) for structured speech, and Spontaneous Speech‐Based Storytelling (Task 3) for semi‐structured speech (Figure 1). We collected 129 speech samples from 21 participants. Using Recursive Feature Elimination with Cross‐Validation (RFECV), we selected 32 key features from over 1,700 acoustic and linguistic ones for classification (Figure 2). We framed the evaluation process as a combinatorial problem where y=f(Task1 Task 2 Task 3). Here, y indicates whether a person is at risk of MCI, and f represents the predictive model that we trained with the speech samples.ResultThe decoding analysis revealed that the combination of Task_1, 2 and 3 achieved the highest AUC performance (AUC 0.963; 100% ratio). Relative to this maximum performance, the combination of Task_2 + Task_3 achieved 0.869 (90.2%), followed by Task_3 0.822 (85.4%), and Task_1 + Task_3 0.817 (84.8%; Figure 3A). Among these, using all tasks resulted in the best classificationResultsan AUC of 0.963, specificity of 0.633, and sensitivity of 0.977 (Figure 3B). The feature importance analysis revealed that 1st quantile regression coefficient of MFCC[14] and 50% upper level time of MFCC delta[9] were the most significant features contributing to the classification model (Figure 3C). Lastly, the acoustic features derived from Task_3, pcm_zcr_sma_de_lpc4 (representing the fourth coefficient of linear predictive coding; LPC) and pcm_zcr_sma_de_lpgain (reflecting the signal‐to‐noise ratio based on energy distribution), showed clear differences between normal and patient groups (Figure 3D).ConclusionAmong various task combinations, the combination of Task_1+Task_2+Task_3 consistently achieved the best results, with Task 3 being included in all high‐performance combinations. Feature importance analysis and target variable distribution further emphasized the greater contribution of acoustic features compared to linguistic features in classification performance.

Delay pervades humanmachine systemsfrom synaptic transmission and axonal conduction to sensing, computation, and actuation. Traditionally viewed as a nuisance to minimize, latency can instead serve as a functional design parameter when it is internal, structured, and predictively compensated. Theoretically, the FLW connects timing-dependent plasticity and predictive coding in neural systems to an analysis of stability in delay-differential equations which shows that moderate delays with structure can stabilize dynamics rather than break them. Practically, this framework informs delay-aware design in brain-computer interfaces and rehabilitation robotics wherein user-experienced delay has to be minimized so as to retain implicit learning and agency. In contrast, internal controller delay can be structured to enhance stability and adaptation. By integrating neuroscience, psychophysics, and control theory, this review proposes delay as both a constraint and a resource, offering a biologically grounded principle for building robust, human-compatible feedback systems.

The hippocampus is classically linked to memory, yet increasing evidence points to a broader role in perceptual inference and deviance detection. Predictive coding theories propose that perception minimizes mismatches between expected and actual sensory input, expressed in neural signatures such as mismatch negativity (MMN) and P300. Although MMN arises mainly from sensory and prefrontal cortices, the hippocampus is anatomically interconnected with both and may also contribute to prediction error processing. We recorded single- and multi-unit activity and local field potentials (LFPs) from DG and CA1 in urethane-anesthetized rats during an auditory oddball paradigm and a no-repetition control sequence to dissociate prediction error from repetition suppression. Approximately 20% of hippocampal neurons were sound responsive, and a subset showed deviant selectivity. Spiking activity predominantly reflected prediction errors, while LFPs revealed complementary contributions from repetition suppression and prediction error. Early LFP components were enhanced for randomly presented deviants, whereas later components within the P300 latency range were stronger for predictable deviants, indicating temporally distinct phases of error signaling and top-down modulation. These findings identify the hippocampus as an active contributor to auditory deviance detection and support a hierarchical model in which hippocampal circuits participate in predictive sensory processing beyond memory.

This perspective paper explores the bidirectional influence between language emergence and the relational structure of subjective experiences, termed qualia structure, and lays out a constructive approach to the intricate dependency between the two. We hypothesize that the emergence of languages with distributional semantics (e.g., syntactic-semantic structures) is linked to the coordination of internal representations shaped by experience, potentially facilitating more structured language through reciprocal influence. This hypothesized mutual dependency connects to recent advancements in AI and symbol emergence robotics, and is explored within this paper through theoretical frameworks such as the collective predictive coding. Computational studies show that neural network-based language models form systematically structured internal representations, and multimodal language models can share representations between language and perceptual information. This perspective suggests that language emergence serves not only as a mechanism for creating a communication tool but also as a mechanism for allowing people to realize shared understanding of qualitative experiences. The paper discusses the implications of this bidirectional influence in the context of consciousness studies, linguistics, and cognitive science, and outlines future constructive research directions to further explore this dynamic relationship between language emergence and qualia structure.

The concept of timing is an interesting way to understand how the body and brain construct the concept of self, but also how self-distortions arise in the case of psychosis. Analysis of temporal representations in psychosis highlights a deficit that includes both the subjective experience of the flow of time, i.e., time perception, and the ability to process temporal information inherent to any perceptual event, i.e., perceptual timing. The representation of the self is stabilized within temporal windows, and thus the self is experienced as continuous in time. Disturbance in the sense of time, in the form of a loss of temporal continuity, has been described by phenomenologists as a central subjective experience of schizophrenia. The positive symptoms of schizophrenia are associated with overestimation of interval timing, i.e., an acceleration of the 'internal clock', while dopamine neurotransmission is likely to regulate the speed of the internal clock. Moreover, findings highlight the importance of interoceptive precision as an aspect of time perception, since accuracy in time perception is related to interoceptive accuracy and vagal activity. Insula contributes significantly to the total awareness of reality. Global emotional moments and meta-representations of the conscious self are created in the anterior insula. In psychosis, the interaction between the default-mode network and the frontoparietal executive network is disrupted by aberrant salience signals from the right anterior insula. Here, we describe the role of the insula as a key hub for the recognition of major aspects of the self, in parallel with the role of interoceptive predictive coding, which reflects the contribution of the insula to the temporality of the self. Based on the above, new insights focus on the development and implementation of rehabilitation strategies that specifically target the temporal deficits observed in psychosis. New therapeutic interventions are based on sensory education and enhancing the multisensory integration of these patients.

Artificial neural networks with backpropagation algorithms have successfully simulated the hierarchical biological brain structures to realize machine learning and describe biological brain learning in many cognitive tasks. However, backpropagation algorithms do not fully follow the rules of biological brain learning to update weights and transmit information, which affects the biological plausibility of backpropagation in the field of neuroscience. The objective of this review is to investigate predictive coding theory as a biologically plausibility alternative to backpropagation, examining both its theoretical potential and practical application feasibility. The predictive coding proposes that the brain minimizes the error between external inputs and expectations by continuously generating and updating internal predictive models, thereby efficiently understanding and interpreting sensory information. Compared with backpropagation, the two core advantages of predictive coding lie in its capacity for efficient local computation and inherent biological plausibility, making it a promising alternative approach to backpropagation. A series of neuroscience experiments have further validated the role of predictive coding in perception, motor control, and cognitive function, highlighting its significance in the study of brain credit assignment mechanisms.

Recurrent neural networks (RNNs) have emerged as a prominent tool for modeling cortical function. However, their conventional architecture is fundamentally lacking in physiological and anatomical fidelity, often raising questions regarding the validity of the insights gleaned from them. Our work therefore develops mathematically grounded methods that let us simultaneously incorporate Dale’s law with highly sparse connectivity motifs into the RNN training pipeline such that the performance of our constrained models empirically matches that of RNNs trained without any constraints. We subsequently demonstrate the utility of our methods for inferring multi-regional interactions by training RNN models with data-driven, cell type–specific connectivity constraints to reconstruct two-photon calcium imaging data during visual behavior in mice spread across multiple cortical layers and brain areas. The interactions inferred by our models corroborate experimental findings in agreement with the theory of predictive coding, across both long and short timescales.

Human auditory cognition involves multiregional hierarchical processing, yet the neural mechanisms integrating cortical layers and large-scale networks remain unclear. Using high-resolution fMRI, we investigate cross-spatial scale dynamics during hierarchical auditory processing. Distinct activations across layers of superior temporal gyrus (STG) and inferior frontal gyrus (IFG) are detected during standard stimuli and violation. We identify the layer-specific effective connectivity among key nodes in the auditory hierarchy using dynamic causal modelling. The results are consistent with hierarchical predictive coding schemes: (i) tone and sequence processing provide input to the superficial and middle layers of STG, respectively. (ii) Forward connections came from the superficial layers of STG, while (iii) backward connections implicated the deep layers of IFG. The integration of mesoscale and macroscale activities confirms the interaction of superficial/middle layers of STG with higher-level regions. These findings elucidate a multiscale mechanism coordinating cortical layers and distributed networks in hierarchical auditory processing.

Hierarchical MMN subcomponents in schizophrenia: Predictive coding biomarkers and clinical translation.

Understanding the Charles Bonnet syndrome: An updated review.

Predictive coding in the primate brain: From visual to fronto-limbic systems