Predictive Coding Framework Research Articles

Simulating Developmental and Individual Differences of Drawing Behavior in Children Using a Predictive Coding Model.

Predictive coding has recently been proposed as a mechanistic approach to explain human perception and behavior based on the integration of perceptual stimuli (bottom-up information) and the predictions about the world based on previous experience (top-down information). However, the gap between the computational accounts of cognition and evidence of behavioral studies remains large. In this study, we used a computational model of drawing based on the mechanisms of predictive coding to systematically investigate the effects of the precision of top-down and bottom-up information when performing a drawing completion task. The results indicated that sufficient precision of both signals was required for the successful completion of the stimuli and that a reduced precision in either sensory or prediction (i.e., prior) information led to different types of atypical drawing behavior. We compared the drawings produced by our model to a dataset of drawings created by children aged between 2 and 8 years old who drew on incomplete drawings. This comparison revealed that a gradual increase in children's precision of top-down and bottom-up information as they develop effectively explains the observed change of drawing style from scribbling toward representational drawing. Furthermore, individual differences that are prevalent in children's drawings, might arise from different developmental pathways regarding the precision of these two signals. Based on these findings we propose a theory of how both general and individual development of drawing could be explained in a unified manner within the framework of predictive coding.

A Predictive Processing Model of Episodic Memory and Time Perception

Human perception and experience of time are strongly influenced by ongoing stimulation, memory of past experiences, and required task context. When paying attention to time, time experience seems to expand; when distracted, it seems to contract. When considering time based on memory, the experience may be different than what is in the moment, exemplified by sayings like "time flies when you're having fun." Experience of time also depends on the content of perceptual experience-rapidly changing or complex perceptual scenes seem longer in duration than less dynamic ones. The complexity of interactions among attention, memory, and perceptual stimulation is a likely reason that an overarching theory of time perception has been difficult to achieve. Here, we introduce a model of perceptual processing and episodic memory that makes use of hierarchical predictive coding, short-term plasticity, spatiotemporal attention, and episodic memory formation and recall, and apply this model to the problem of human time perception. In an experiment with approximately 13,000 human participants, we investigated the effects of memory, cognitive load, and stimulus content on duration reports of dynamic natural scenes up to about 1 minute long. Using our model to generate duration estimates, we compared human and model performance. Model-based estimates replicated key qualitative biases, including differences by cognitive load (attention), scene type (stimulation), and whether the judgment was made based on current or remembered experience (memory). Our work provides a comprehensive model of human time perception and a foundation for exploring the computational basis of episodic memory within a hierarchical predictive coding framework.

Emergence of associative learning in a neuromorphic inference network

Objective. In the theoretical framework of predictive coding and active inference, the brain can be viewed as instantiating a rich generative model of the world that predicts incoming sensory data while continuously updating its parameters via minimization of prediction errors. While this theory has been successfully applied to cognitive processes—by modelling the activity of functional neural networks at a mesoscopic scale—the validity of the approach when modelling neurons as an ensemble of inferring agents, in a biologically plausible architecture, remained to be explored. Approach. We modelled a simplified cerebellar circuit with individual neurons acting as Bayesian agents to simulate the classical delayed eyeblink conditioning protocol. Neurons and synapses adjusted their activity to minimize their prediction error, which was used as the network cost function. This cerebellar network was then implemented in hardware by replicating digital neuronal elements via a low-power microcontroller. Main results. Persistent changes of synaptic strength—that mirrored neurophysiological observations—emerged via local (neurocentric) prediction error minimization, leading to the expression of associative learning. The same paradigm was effectively emulated in low-power hardware showing remarkably efficient performance compared to conventional neuromorphic architectures. Significance. These findings show that: (a) an ensemble of free energy minimizing neurons—organized in a biological plausible architecture—can recapitulate functional self-organization observed in nature, such as associative plasticity, and (b) a neuromorphic network of inference units can learn unsupervised tasks without embedding predefined learning rules in the circuit, thus providing a potential avenue to a novel form of brain-inspired artificial intelligence.

A Predictive Coding Framework for Understanding Major Depression.

Predictive coding models of brain processing propose that top-down cortical signals promote efficient neural signaling by carrying predictions about incoming sensory information. These “priors” serve to constrain bottom-up signal propagation where prediction errors are carried via feedforward mechanisms. Depression, traditionally viewed as a disorder characterized by negative cognitive biases, is associated with disrupted reward prediction error encoding and signaling. Accumulating evidence also suggests that depression is characterized by impaired local and long-range prediction signaling across multiple sensory domains. This review highlights the electrophysiological and neuroimaging evidence for disrupted predictive processing in depression. The discussion is framed around the manner in which disrupted generative predictions about the sensorium could lead to depressive symptomatology, including anhedonia and negative bias. In particular, the review focuses on studies of sensory deviance detection and reward processing, highlighting research evidence for both disrupted generative predictions and prediction error signaling in depression. The role of the monoaminergic and glutamatergic systems in predictive coding processes is also discussed. This review provides a novel framework for understanding depression using predictive coding principles and establishes a foundational roadmap for potential future research.

Constructing the hierarchy of predictive auditory sequences in the marmoset brain.

Our brains constantly generate predictions of sensory input that are compared with actual inputs, propagate the prediction-errors through a hierarchy of brain regions, and subsequently update the internal predictions of the world. However, the essential feature of predictive coding, the notion of hierarchical depth and its neural mechanisms, remains largely unexplored. Here, we investigated the hierarchical depth of predictive auditory processing by combining functional magnetic resonance imaging (fMRI) and high-density whole-brain electrocorticography (ECoG) in marmoset monkeys during an auditory local-global paradigm in which the temporal regularities of the stimuli were designed at two hierarchical levels. The prediction-errors and prediction updates were examined as neural responses to auditory mismatches and omissions. Using fMRI, we identified a hierarchical gradient along the auditory pathway: midbrain and sensory regions represented local, shorter-time-scale predictive processing followed by associative auditory regions, whereas anterior temporal and prefrontal areas represented global, longer-time-scale sequence processing. The complementary ECoG recordings confirmed the activations at cortical surface areas and further differentiated the signals of prediction-error and update, which were transmitted via putative bottom-up γ and top-down β oscillations, respectively. Furthermore, omission responses caused by absence of input, reflecting solely the two levels of prediction signals that are unique to the hierarchical predictive coding framework, demonstrated the hierarchical top-down process of predictions in the auditory, temporal, and prefrontal areas. Thus, our findings support the hierarchical predictive coding framework, and outline how neural networks and spatiotemporal dynamics are used to represent and arrange a hierarchical structure of auditory sequences in the marmoset brain.

Actions of a shaken heart: Interoception interacts with action processing

In the present study, we investigated the modulatory influence of the unconscious, bodily arousal on motor-related embodied information. Specifically, we examined how the interoceptive prediction error interacts with the event-related potentials linked to action-effect processing. Participants were asked to perform a task with self-initiated or externally-triggered sounds while receiving synchronous or false auditory cardiac feedback. The results found that interaction of interoceptive manipulation and action-effect processing modulates the frontal subcomponent of the P3 response. During the synchronous cardiac feedback, the P3 response to self-initiated tones was enhanced. During the false cardiac feedback, the frontal cortical response was reversed. N1 and P2 components were affected by the interoceptive manipulation, but not by the interaction of interoception and action processing. These findings provide experimental support for the theoretical accounts of the interaction between interoception and action processing within a framework of predictive coding, manifested particularly in the higher stages of action processing.

Unsupervised Speech Segmentation and Variable Rate Representation Learning Using Segmental Contrastive Predictive Coding

Typically, unsupervised segmentation of speech into the phone- and word-like units are treated as separate tasks and are often done via different methods which do not fully leverage the inter-dependence of the two tasks. Here, we unify them and propose a technique that can jointly perform both, showing that these two tasks indeed benefit from each other. Recent attempts employ self-supervised learning, such as contrastive predictive coding (CPC), where the next frame is predicted given past context. However, CPC only looks at the audio signal’s frame-level structure. We overcome this limitation with a segmental contrastive predictive coding (SCPC) framework to model the signal structure at a higher level, e.g., phone level. A convolutional neural network learns frame-level representation from the raw waveform via noise-contrastive estimation (NCE). A differentiable boundary detector finds variable-length segments, which are then used to optimize a segment encoder via NCE to learn segment representations. The differentiable boundary detector allows us to train frame-level and segment-level encoders jointly. Experiments show that our single model outperforms existing phone and word segmentation methods on TIMIT and Buckeye datasets. We analyze the impact of the threshold on boundary detector performance, and our results suggest that automatically learning the boundary threshold can be as effective as manually tuning that threshold. We discover that phone class impacts the boundary detection performance, and the boundaries between successive vowels or semivowels are the most difficult. Finally, we use SCPC to extract speech features at the segment level rather than at the uniformly spaced frame level (e.g., 10 ms) and produce variable rate representations that change according to the contents of the utterance. We can lower the feature extraction rate from the typical 100 Hz to as low as 14.5 Hz on average while still outperforming the hand-crafted features such as MFCC on the linear phone classification task.

Sensory Attenuation in the Auditory Modality as a Window Into Predictive Processing

Self-generated auditory input is perceived less loudly than the same sounds generated externally. The existence of this phenomenon, called Sensory Attenuation (SA), has been studied for decades and is often explained by motor-based forward models. Recent developments in the research of SA, however, challenge these models. We review the current state of knowledge regarding theoretical implications about the significance of Sensory Attenuation and its role in human behavior and functioning. Focusing on behavioral and electrophysiological results in the auditory domain, we provide an overview of the characteristics and limitations of existing SA paradigms and highlight the problem of isolating SA from other predictive mechanisms. Finally, we explore different hypotheses attempting to explain heterogeneous empirical findings, and the impact of the Predictive Coding Framework in this research area.

Circular Causality of Emotions in Moving Pictures

Abstract In the framework of predictive coding, as explained by Giovanni Pezzulo in his article Why do you fear the bogeyman? An embodied predictive coding model of perceptual inference (2014), humans construct instances of emotions by a double arrow of explanation of stimuli. Top-down cognitive models explain in a predictive fashion the emotional value of stimuli. At the same time, feelings and emotions depend on the perception of internal changes in the body. When confronted with uncertain auditory and visual information, a multimodal internal state assigns more weight to interoceptive information (rather than auditory and visual information) like visceral and autonomic states as hunger or thirst (motivational conditions). In short, an emotional mood can constrain the construction of a particular instance of emotion. This observation suggests that the dynamics of generative processes of Bayesian inference contain a mechanism of bidirectional link between perceptual and cognitive inference and feelings and emotions. In other words, “subjective feeling states and emotions influence perceptual and cognitive inference, which in turn produce new subjective feeling states and emotions” as a self-fulfilling prophecy (Pezzulo 2014, 908). This article focuses on the short introductory scene from Steven Spielberg’s Jaws (1975), claiming that the construction / emergence of the fear and sadness emotions are created out of the circular causal coupling instantiated between cinematic bottom-up mood cues and top-down cognitive explanations.

Neural correlates of interoceptive accuracy: Beyond cardioception.

Interoceptive accuracy (IAc), the precision with which one assesses the signals arising from one's own body, is receiving increasing attention in the literature. IAc has mainly been approached as an individual trait and has been investigated through the cardiac modality using mostly non-ecological methods. Such studies consensually designate the anterior insular cortex as the main brain correlate of IAc. However, there is a lack of brain imaging studies investigating IAc in a broader and more ecological way. Here, we used a novel ecological task in which participants monitored their general bodily reactions to external events and investigated brain regions subtending intraindividual (i.e. trial-by-trial) variations of IAc. At each trial, participants had to rate the intensity of their bodily reactions to an emotional picture. We recorded participants' skin conductance response (SCR) to the picture as an indicator of actual physiological response intensity. We fitted a functional magnetic resonance imaging (fMRI) model using, as regressors, the SCR value, the rating and the product of the two (as a proxy of participants' IAc) obtained trial per trial. We observed that activity in the dorsomedial prefrontal cortex (dmPFC) increased when individuals' IAc decreased. This result reveals general mechanism of error processing in intraindividual variations of IAc, which are unspecific to interoception. Our result has a practical impact in the clinical domain. Namely, it supports the predictive coding framework whereby IAc deficits may reflect impairments in processing a mismatch between actual interoceptive signals and predictions.

Live agent preference and social action monitoring in the macaque mid-superior temporal sulcus region

Mentalizing, the ability to infer the mental states of others, is a cornerstone of adaptive social intelligence. While functional brain mapping of human mentalizing has progressed considerably, its evolutionary signature in nonhuman primates remains debated. The discovery that the middle part of the macaque superior temporal sulcus (mid-STS) region has a connectional fingerprint most similar to the human temporoparietal junction (TPJ)-a crucial node in the mentalizing network-raises the possibility that these cortical areas may also share basic functional properties associated with mentalizing. Here, we show that this is the case in aspects of a preference for live social interactions and in a theoretical framework of predictive coding. Macaque monkeys were trained to perform a turn-taking choice task with another real monkey partner sitting directly face-to-face or a filmed partner appearing in prerecorded videos. We found that about three-fourths of task-related mid-STS neurons exhibited agent-dependent activity, most responding selectively or preferentially to the partner's action. At the population level, activities of these partner-type neurons were significantly greater under live-partner compared to video-recorded-partner task conditions. Furthermore, a subset of the partner-type neurons responded proactively when predictions about the partner's action were violated. This prediction error coding was specific to the action domain; almost none of the neurons signaled error in the prediction of reward. The present findings highlight unique roles of the macaque mid-STS at the single-neuron level and further delineate its functional parallels with the human TPJ in social cognitive processes associated with mentalizing.

Social prediction in pediatric patients with congenital, non-progressive malformations of the cerebellum: From deficits in predicting movements to rehabilitation in virtual reality

It has been proposed that impairments of the predictive function exerted by the cerebellum may account for social cognition deficits. Here, we integrated cerebellar functions in a predictive coding framework to elucidate how congenital, non-progressive cerebellar alterations could affect the predictive processing of others' behavior. Experiment 1 demonstrated that cerebellar patients were impaired in relying on contextual information during prediction of other persons' movement, and this impairment was significantly associated with social cognition abilities. Experiment 2 indicated that children and adolescents with congenital, non-progressive cerebellar malformation showed a domain-general deficit in using contextual information to predict both others' movements and physical events, and that this impairment was independent from patients' cognitive abilities. Experiment 3 provided first evidence that a social-prediction training in virtual reality could boost the ability to use context-based predictions to understand others’ intentions. These findings shed new light on the predictive role of the cerebellum and its contribution to social cognition, paving the way for new approaches to the rehabilitation of the Cerebellar Cognitive Affective Syndrome.

Bidirectional interaction between visual and motor generative models using Predictive Coding and Active Inference

In this work, we build upon the Active Inference (AIF) and Predictive Coding (PC) frameworks to propose a neural architecture comprising a generative model for sensory prediction, and a distinct generative model for motor trajectories. We highlight how sequences of sensory predictions can act as rails guiding learning, control and online adaptation of motor trajectories. We furthermore inquire the effects of bidirectional interactions between the motor and the visual modules. The architecture is tested on the control of a simulated robotic arm learning to reproduce handwritten letters.

Contrastive Predictive Coding for Human Activity Recognition

Feature extraction is crucial for human activity recognition (HAR) using body-worn movement sensors. Recently, learned representations have been used successfully, offering promising alternatives to manually engineered features. Our work focuses on effective use of small amounts of labeled data and the opportunistic exploitation of unlabeled data that are straightforward to collect in mobile and ubiquitous computing scenarios. We hypothesize and demonstrate that explicitly considering the temporality of sensor data at representation level plays an important role for effective HAR in challenging scenarios. We introduce the Contrastive Predictive Coding (CPC) framework to human activity recognition, which captures the temporal structure of sensor data streams. Through a range of experimental evaluations on real-life recognition tasks, we demonstrate its effectiveness for improved HAR. CPC-based pre-training is self-supervised, and the resulting learned representations can be integrated into standard activity chains. It leads to significantly improved recognition performance when only small amounts of labeled training data are available, thereby demonstrating the practical value of our approach. Through a series of experiments, we also develop guidelines to help practitioners adapt and modify the framework towards other mobile and ubiquitous computing scenarios.

Can expectation suppression be explained by reduced attention to predictable stimuli?

The expectation-suppression effect – reduced stimulus-evoked responses to expected stimuli – is widely considered to be an empirical hallmark of reduced prediction errors in the framework of predictive coding. Here we challenge this notion by proposing that that expectation suppression could be explained by a reduced attention effect. Specifically, we argue that reduced responses to predictable stimuli can also be explained by a reduced saliency-driven allocation of attention. We base our discussion mainly on findings in the visual cortex and propose that resolving this controversy requires the assessment of qualitative differences between the ways in which attention and surprise enhance brain responses.

Interactions of Medial and Lateral Prefrontal Cortex in Hierarchical Predictive Coding.

Cognitive control and decision-making rely on the interplay of medial and lateral prefrontal cortex (mPFC/lPFC), particularly for circumstances in which correct behavior requires integrating and selecting among multiple sources of interrelated information. While the interaction between mPFC and lPFC is generally acknowledged as a crucial circuit in adaptive behavior, the nature of this interaction remains open to debate, with various proposals suggesting complementary roles in (i) signaling the need for and implementing control, (ii) identifying and selecting appropriate behavioral policies from a candidate set, and (iii) constructing behavioral schemata for performance of structured tasks. Although these proposed roles capture salient aspects of conjoint mPFC/lPFC function, none are sufficiently well-specified to provide a detailed account of the continuous interaction of the two regions during ongoing behavior. A recent computational model of mPFC and lPFC, the Hierarchical Error Representation (HER) model, places the regions within the framework of hierarchical predictive coding, and suggests how they interact during behavioral periods preceding and following salient events. In this manuscript, we extend the HER model to incorporate real-time temporal dynamics and demonstrate how the extended model is able to capture single-unit neurophysiological, behavioral, and network effects previously reported in the literature. Our results add to the wide range of results that can be accounted for by the HER model, and provide further evidence for predictive coding as a unifying framework for understanding PFC function and organization.

A continuum hypothesis of psychotomimetic rapid antidepressants.

Ketamine, classical psychedelics and sleep deprivation are associated with rapid effects on depression. Interestingly, these interventions also have common psychotomimetic actions, mirroring aspects of psychosis such as an altered sense of self, perceptual distortions and distorted thinking. This raises the question whether these interventions might be acute antidepressants through the same mechanisms that underlie some of their psychotomimetic effects. That is, perhaps some symptoms of depression can be understood as occupying the opposite end of a spectrum where elements of psychosis can be found on the other side. This review aims at reviewing the evidence underlying a proposed continuum hypothesis of psychotomimetic rapid antidepressants, suggesting that a range of psychotomimetic interventions are also acute antidepressants as well as trying to explain these common features in a hierarchical predictive coding framework, where we hypothesise that these interventions share a common mechanism by increasing the flexibility of prior expectations. Neurobiological mechanisms at play and the role of different neuromodulatory systems affected by these interventions and their role in controlling the precision of prior expectations and new sensory evidence will be reviewed. The proposed hypothesis will also be discussed in relation to other existing theories of antidepressants. We also suggest a number of novel experiments to test the hypothesis and highlight research areas that could provide further insights, in the hope to better understand the acute antidepressant properties of these interventions.

The role of isochrony in speech perception in noise

The role of isochrony in speech—the hypothetical division of speech units into equal duration intervals—has been the subject of a long-standing debate. Current approaches in neurosciences have brought new perspectives in that debate through the theoretical framework of predictive coding and cortical oscillations. Here we assess the comparative roles of naturalness and isochrony in the intelligibility of speech in noise for French and English, two languages representative of two well-established contrastive rhythm classes. We show that both top-down predictions associated with the natural timing of speech and to a lesser extent bottom-up predictions associated with isochrony at a syllabic timescale improve intelligibility. We found a similar pattern of results for both languages, suggesting that temporal characterisation of speech from different rhythm classes could be unified around a single core speech unit, with neurophysiologically defined duration and linguistically anchored temporal location. Taken together, our results suggest that isochrony does not seem to be a main dimension of speech processing, but may be a consequence of neurobiological processing constraints, manifesting in behavioural performance and ultimately explaining why isochronous stimuli occupy a particular status in speech and human perception in general.

Rapid computations of spectrotemporal prediction error support perception of degraded speech.

Human speech perception can be described as Bayesian perceptual inference but how are these Bayesian computations instantiated neurally? We used magnetoencephalographic recordings of brain responses to degraded spoken words and experimentally manipulated signal quality and prior knowledge. We first demonstrate that spectrotemporal modulations in speech are more strongly represented in neural responses than alternative speech representations (e.g. spectrogram or articulatory features). Critically, we found an interaction between speech signal quality and expectations from prior written text on the quality of neural representations; increased signal quality enhanced neural representations of speech that mismatched with prior expectations, but led to greater suppression of speech that matched prior expectations. This interaction is a unique neural signature of prediction error computations and is apparent in neural responses within 100 ms of speech input. Our findings contribute to the detailed specification of a computational model of speech perception based on predictive coding frameworks.

Auditory Mismatch Negativity Under Predictive Coding Framework and Its Role in Psychotic Disorders.

Traditional neuroscience sees sensory perception as a simple feedforward process. This view is challenged by the predictive coding model in recent years due to the robust evidence researchers had found on how our prediction could influence perception. In the first half of this article, we reviewed the concept of predictive brain and some empirical evidence of sensory prediction in visual and auditory processing. The predictive function along the auditory pathway was mainly studied by mismatch negativity (MMN)—a brain response to an unexpected disruption of regularity. We summarized a range of MMN paradigms and discussed how they could contribute to the theoretical development of the predictive coding neural network by the mechanism of adaptation and deviance detection. Such methodological and conceptual evolution sharpen MMN as a tool to better understand the structural and functional brain abnormality for neuropsychiatric disorder such as schizophrenia.