Perceptual inference (2023)

Introduction

Subsequent to Barlow's proposal that perceptual awareness is linked to the firing of individual neurons (Barlow, 1961, Barlow, 1972), visual perception had been regarded as the ultimate outcome of sensory input through a feedforward process arising in the retina. The simplest impression of such a process came from the pioneering description of simple, complex and hypercomplex cells in areas 17, 18 and 19 of the cat neocortex (Hubel and Wiesel, 1962, Hubel and Wiesel, 1965) indicating a progressive feedforward increase in the complexity of physiological receptive field properties with successive hierarchical stages (Felleman and van Essen, 1991). Nevertheless in a different context, when behaviour is also under consideration, the input from sensory organs has been traditionally seen not as feedforward but as a feedback input (e.g. Eldred and Buchwald, 1967, af Klint et al., 2010).

An approach to visual perception that bridges this apparent divide, proposed indeed more than a century ago by Helmut von Helmholtz (1910), puts an emphasis on the formation of a percept within a process of evaluation. On Helmholtz's suggestion, the evaluation involves a test of a hypothesis about what is being seen based on “inductive inferences” gained from “sensations”. By inductive inference Helmholtz meant that perceptions are conclusions based not only on present sensations but also with reference to past sensations of the objects perceived. Latent within this conceptualisation is the idea that the perceived image is at least partly the outcome of stored information – a stored representation, that is a memory – of that object or of similar objects in similar contexts. This was potentially the first proposal of a top-down influence in perception. It regards perception not primarily as a sensory phenomenon but as perceptual inference relying on internal models built through past experience. Helmholtz's idea of perceptual inference has been revived by computational models of perception relying on statistical inference (Hinton and Sejnowski, 1983, Dayan et al., 1995).

Between Helmholtz's proposal and its recent revival, several other theories of perception were advanced. Most of these proposals were based on the hierarchical organisation of the visual system and the increasing complexity of visual fields in occipital cortex in particular, and were influenced by the binding theory.

The binding theory in linguistics is a theory of syntax and phrase structure grammar, given that the same words in a sentence when rearranged can have different meanings (e.g. “Mark said he was present” vs “He said Mark was present”). The idea was subsequently applied to visual perception as a model solution to the “binding problem”. The binding problem considers that features of an object need to be bound together by some neuronal mechanism across a population of neurons, so that the object can be perceived as a whole. It rather assumes that representations of that object are not possible unless a bottom-up binding of its features can bring it into consciousness and make it a subject of attention.

A further idea proposes that the mechanism by which the features of objects are bound together is spike-correlations (von der Malsburg, 1981, von der Malsburg and Schneider, 1986, Iwabuchi, 1998). This idea assumes that neuronal responses to features of an object need to be bound to a single entity, which is ultimately achieved through spike synchrony. Certain predictions from this theory have not found experimental support (Golledge et al., 2003, Thiele and Stoner, 2003, Rolls et al., 2003a, Rolls et al., 2003b, Nielsen et al., 2006, Dong et al., 2008). One difficult problem arising, if objects had no permanent representation in the brain but had to be bound anew every time, from scratch so to speak, through an essentially bottom-up binding of their features, would be how would then an object be associated through experience with another object, or with a place or with an action or conditioned response.

Two alternative theories, those of feature integration and the structural description theories, place a greater emphasis on lateral or top-down influences.

The feature integration theory, developed by Treisman and Gelade, 1980, Treisman, 1985, Treisman, 1986 proposes that different attentional mechanisms are responsible for binding different features into consciously experienced wholes. The theory has been one of the most influential psychological models of human visual attention.

According to Treisman, in a first step to visual processing, several primary visual features are processed and represented with separate feature maps that are later integrated into a “saliency map” that can be accessed in order to direct attention to the most conspicuous areas. Although this model does not preclude and is not aimed at precluding top-down influences, the modularity of this system based on “features” and its ability to function at the pre-attentive state, puts an emphasis on the ability to perceive objects on the basis of a bottom-up mechanism arriving stepwise at higher-order features, albeit stored (i.e. already represented internally) in saliency maps. It has been argued, moreover, that bottom-up salience-driven mechanisms can draw attentional selection only in as much as there exist learned associations between the relevant stimuli and rewards (Anderson, 2013).

The structural description theories generally propose that complex shapes are built up of more elementary features or primitives. In these sets of theories (reviewed by Peissig and Tarr, 2007), an actual object is thought to be represented in the brain by a structural description of its parts, in other words by its features and a syntax that describes how these features are combined. The features can be elementary 3-D primitives as in the theory of Marr and Nishihara (1978) or elementary geometrical shapes (Pentland, 1984) called geons in the Recognition by Components (RBC) theory of Biederman (1987). In the latter theory these elementary shapes behave as phonemes in speech, and are derived from contrasts of properties of edges in a two-dimensional image (curvature, collinearity, symmetry, parallelism and cotermination). These theories were aiming to explain the phenomenon of view invariance, that is the ability to identify an object regardless of the position of the viewer relative to the position of the object. Although some researchers have used the RBC theory to some degree successfully in describing neuronal responses to simple artificial combinations of geometrical features (e.g. Kayaert et al., 2003), experimental evidence has been equivocal when neuronal responses to actual and more complex natural objects (e.g. animals) are being considered (e.g. Wachsmuth et al., 1994). Single neurons often respond with view-dependent preferences. Moreover, the discovery that IT neurons tend to prefer similar overall shapes regardless of local differences, including shapes where there is an aspect ratio or orientation preference (Baldassi et al., 2013), is not in agreement with the RBC theory.

Observations that neuronal responses are tuned to increasingly more complex visual stimuli over successive hierarchical levels (Hubel and Wiesel, 1962, Hubel and Wiesel, 1965, Gross et al., 1972, Rolls et al., 1977, Bruce et al., 1981, Kobatake and Tanaka, 1994), have given rise to hierarchical models of invariant face and object recognition such as those of the hierarchical GRBF model of Poggio and Edelman (1990), its development by Riesenhuber and Poggio (1999), the Perrett and Oram (1998) model of visual recognition and the VisNet model of Wallis and Rolls (1997).

A strictly hierarchical conception focusing on a stagewise bottom-up creation of percepts would predict that attention is the result of perception and that it is drawn to a stimulus after the stimulus has been perceived. This would lead to the question of what would unify the stimulus so that attention can be directed to it, the so-called binding problem. Some theories of attention indeed put the direct perception of stimulus before the attention to it (e.g. Gregory, 1980, Theeuwes, 1993, Theeuwes, 1994, Simons, 2000). The hierarchical models of invariant face and object recognition rely on model networks being trained with repeated presentations of a set of stimuli, so that in their fully trained state they work as models of the visual system essentially because of their stored memories. Stored memories assume the role of forward models in predicting sensory input, obviating the need for a bottom-up binding of its features.

Section snippets

Perception as a top-down guided process

That stored memories are used in perception is hardly a new idea. The idea had been expressed in 1890 by William James: “Whilst part of what we perceive comes from the object before us, another part (and it may be the larger part) always comes out of our own head”, a statement sometimes referred to as William James’ law of perception.

The cortical stimulation studies of Wilder Penfield (e.g. Penfield, 1961) had also suggested that the cortex is by itself able to produce percepts on the basis of

Prerequisites for perceptual inference

Potential perceptual inference correlates are approximately equivalent to a priming suprathreshold activity at the executive level in anticipation of a stimulus, such as during working memory, correlating with partially or primarily non-spiking changes of neural networks at the sensory level, prior to the direct presentation of an actual stimulus. Among such correlates might be (1) sub-threshold neuronal facilitation in the absence of an overt stimulus, (2) change in synaptic efficacy in

Anatomical arrangement of the network for perceptual inference

The cerebellar (Miall et al., 1993) and cerebral (Pruszynski et al., 2011) cortices are conceptualised as influencing via a polysynaptic pathway a sensorimotor reflex action. In the vestibulo-ocular reflex for example (Fig. 2), the cerebellum and associated structures such as the vestibular nucleus are seen as playing the role of the comparator in a circuit that resembles a servomechanism (Gonshor and Melvill Jones, 1976b). In more recent conceptualisations, the standard model of a

Predictions based on the concept of perceptual inference.

Several predictions can be made on the basis of perceptual inference. Percepts should be under appropriate conditions generated equally vividly without sensory input. Several cases have been already presented where this can be observed: dreaming, hallucinations and evoked responses to electrical stimulation of the neocortex.

Another expectation compatible with the concept of perceptual inference would be that humans and animals should show increased interest, e.g. preferential looking, to novel

(Video) Bayesian Brain

Conclusion

In conclusion, perception can be understood as a process of evaluation of the sensory input against stored information (memories or forward models) in a process of adaptation of behaviour to the relevant stimuli. New information may be acquired by the organism so as to minimise prediction error. This process is facilitated by attentional selection and the operation of working memory-related activity. The suprathreshold spiking activity at the executive level relating to attention and

Cited by (26)

Perceptual inference, accuracy, and precision in temporal reproduction in schizophrenia
2022, Schizophrenia Research: Cognition
See Also
Perceptual inference
Accumulating evidence suggests that deficits in perceptual inference account for symptoms of schizophrenia. One manifestation of perceptual inference is the central bias, i.e., the tendency to put emphasis on prior experiences over actual events in perceiving incoming sensory stimuli. Using an interval reproduction task, this study aimed to determine whether patients with schizophrenia show a stronger central bias than participants without schizophrenia. In the interval reproduction task, participants were shown a cross on a screen. The cross was replaced with a Gaussian patch for a predetermined time interval, and participants were required to reproduce the interval duration by pressing and releasing the space key. We manipulated the uncertainty of prior information using different interval distributions. We found no difference in the influence of prior information on interval reproduction between patients and controls. However, patients with SZ showed a stronger central bias than healthy participants in the intermediate interval range (approximately 450 ms to 900 ms). It is possible that the patients in SZ have non-uniform deficits associated with interval range or uncertainty of prior information in perceptual inference. Further, the severity of avolition and alogia was correlated with the strength of central bias in SZ. This study provides some insights into the mechanisms underlying the association between schizophrenic symptoms and perceptual inference.
Interoception as modeling, allostasis as control
2022, Biological Psychology
The brain regulates the body by anticipating its needs and attempting to meet them before they arise – a process called allostasis. Allostasis requires a model of the changing sensory conditions within the body, a process called interoception. In this paper, we examine how interoception may provide performance feedback for allostasis. We suggest studying allostasis in terms of control theory, reviewing control theory’s applications to related issues in physiology, motor control, and decision making. We synthesize these by relating them to the important properties of allostatic regulation as a control problem. We then sketch a novel formalism for how the brain might perform allostatic control of the viscera by analogy to skeletomotor control, including a mathematical view on how interoception acts as performance feedback for allostasis. Finally, we suggest ways to test implications of our hypotheses.
Cognitive biases, environmental, patient and personal factors associated with critical care decision making: A scoping review
2021, Journal of Critical Care
Cognitive biases and factors affecting decision making in critical care can potentially lead to life-threatening errors. We aimed to examine the existing evidence on the influence of cognitive biases and factors on decision making in critical care.
(Video) Perceiving is Believing: Crash Course Psychology #7
We conducted a scoping review by searching MEDLINE for articles from 2004 to November 2020. We included studies conducted in physicians that described cognitive biases or factors associated with decision making. During the study process we decided on the method to summarize the evidence, and based on the obtained studies a descriptive summary of findings was the best fit.
Thirty heterogenous studies were included. Four main biases or factors were observed, e.g. cognitive biases, personal factors, environmental factors, and patient factors. Six (20%) studies reported biases associated with decision making comprising omission-, status quo-, implicit-, explicit-, outcome-, and overconfidence bias. Nineteen (63%) studies described personal factors, twenty-two (73%) studies described environmental factors, and sixteen (53%) studies described patient factors.
The current evidence on cognitive biases and factors is heterogenous, but shows they influence clinical decision. Future studies should investigate the prevalence of cognitive biases and factors in clinical practice and their impact on clinical outcomes.
Data gathering ability contributes to visual organization and probabilistic reasoning
2018, Heliyon
Individuals use data gathering methods to inform judgments and behaviors. Effective interaction with the environment depends on these having high accuracy and low noise, but when they become abnormal, aberrant thoughts and perceptions can occur. In this study, we examined if data gathering methods were consistent across tasks that relied on different cognitive abilities, specifically visual perception and probabilistic reasoning. Thirty-four non-clinical participants engaged in the Ebbinghaus Illusion and the Jumping to Conclusions tasks, while also completing questionnaires concerning aspects of delusion formation. A significant, positive correlation was observed between performance on the Ebbinghaus Illusion and the Jumping to Conclusions tasks. Both tasks were negatively correlated with the General Conspiracy Belief Scale. The results suggest an underlying mechanism for data gathering that is consistent across behavioral domains and exists on a continuum in the general population.
Juxtaposing the real-time unfolding of subjective experience and ERP neuromarker dynamics
2017, Consciousness and Cognition
Electroencephalographic (EEG) potentials have remained a valuable source of data and theories concerning neural correlates of consciousness (NCC). The EEG based methods are far from being exhausted and are continually valuable in the quest for the markers of NCC. To set the background for the research presented in this issue, we review the published work on EEG-based markers of NCC. The article is organized according to the principle of the time-course aspect of brain potentials with regard to the stimuli for which subject’s awareness is experimentally measured and/or manipulated. We treat brain potentials as the principal dependent measure as well as independent variable. More specifically, we also draw attention to the fact that in the overwhelming share of studies relative negativization of the ERPs tends to mark NCC.
Bayes’ Theorem in Neurocritical Care: Principles and Practice
2023, Neurocritical Care

View all citing articles on Scopus

Perceptual inference (2023)

Introduction

Section snippets

Perception as a top-down guided process

Prerequisites for perceptual inference

Anatomical arrangement of the network for perceptual inference

Predictions based on the concept of perceptual inference.

Conclusion

Cited by (26)

Perceptual inference, accuracy, and precision in temporal reproduction in schizophrenia

Interoception as modeling, allostasis as control

Cognitive biases, environmental, patient and personal factors associated with critical care decision making: A scoping review

Data gathering ability contributes to visual organization and probabilistic reasoning

Juxtaposing the real-time unfolding of subjective experience and ERP neuromarker dynamics

Bayes’ Theorem in Neurocritical Care: Principles and Practice

Recommended articles (6)

Serotonin mediated immunoregulation and neural functions: Complicity in the aetiology of autism spectrum disorders

Bipolar disorder and antibodies against the N-methyl-d-aspartate receptor: A gate to the involvement of autoimmunity in the pathophysiology of bipolar illness

Cohesion and the aneuploid phenotype in Alzheimer's disease: A tale of genome instability

The muscarinic system, cognition and schizophrenia

Psychiatric disorders and leukocyte telomere length: Underlying mechanisms linking mental illness with cellular aging

Painfree and accurate Bayesian estimation of psychometric functions for (potentially) overdispersed data

Videos