Image interpretation by a single bottom-up top-down cycle

Hierarchical Human Semantic Parsing With Comprehensive Part-Relation Modeling

Modeling the human structure is central for human parsing that extracts pixel-wise semantic information from images. We start with analyzing three types of inference processes over the hierarchical structure of human bodies: direct inference (directly predicting human semantic parts using image information), bottom-up inference (assembling knowledge from constituent parts), and top-down inference (leveraging context from parent nodes). We then formulate the problem as a compositional neural information fusion (CNIF) framework, which assembles the information from the three inference processes in a conditional manner, i.e., considering the confidence of the sources. Based on CNIF, we further present a part-relation-aware human parser (PRHP), which precisely describes three kinds of human part relations, i.e., decomposition, composition, and dependency, by three distinct relation networks. Expressive relation information can be captured by imposing the parameters in the relation networks to satisfy specific geometric characteristics of different relations. By assimilating generic message-passing networks with their edge-typed, convolutional counterparts, PRHP performs iterative reasoning over the human body hierarchy. With these efforts, PRHP provides a more general and powerful form of CNIF, and lays the foundation for more sophisticated and flexible human relation patterns of reasoning. Experiments on five datasets demonstrate that our two human parsers outperform the state-of-the-arts in all cases.

Do cognitive load and ADHD traits affect the tendency to prioritise social information in scenes?

We report two experiments investigating the effect of working memory (WM) load on selective attention. Experiment 1 was a modified version of Lavie et al. and confirmed that increasing memory load disrupted performance in the classic flanker task. Experiment 2 used the same manipulation of WM load to probe attention during the viewing of complex scenes while also investigating individual differences in attention deficit hyperactivity disorder (ADHD) traits. In the image-viewing task, we measured the degree to which fixations targeted each of two crucial objects: (1) a social object (a person in the scene) and (2) a non-social object of higher or lower physical salience. We compared the extent to which increasing WM load would change the pattern of viewing of the physically salient and socially salient objects. If attending to the social item requires greater default voluntary top-down resources, then the viewing of social objects should show stronger modulation by WM load compared with viewing of physically salient objects. The results showed that the social object was fixated to a greater degree than the other object (regardless of physical salience). Increased salience drew fixations away from the background leading to slightly increased fixations on the non-social object, without changing fixations on the social object. Increased levels of ADHD-like traits were associated with fewer fixations on the social object, but only in the high-salient, low-load condition. Importantly, WM load did not affect the number of fixations on the social object. Such findings suggest rather surprisingly that attending to a social area in complex stimuli is not dependent on the availability of voluntary top-down resources.

Face detection in untrained deep neural networks

Face-selective neurons are observed in the primate visual pathway and are considered as the basis of face detection in the brain. However, it has been debated as to whether this neuronal selectivity can arise innately or whether it requires training from visual experience. Here, using a hierarchical deep neural network model of the ventral visual stream, we suggest a mechanism in which face-selectivity arises in the complete absence of training. We found that units selective to faces emerge robustly in randomly initialized networks and that these units reproduce many characteristics observed in monkeys. This innate selectivity also enables the untrained network to perform face-detection tasks. Intriguingly, we observed that units selective to various non-face objects can also arise innately in untrained networks. Our results imply that the random feedforward connections in early, untrained deep neural networks may be sufficient for initializing primitive visual selectivity.

Interpreting Bottom-Up Decision-Making of CNNs via Hierarchical Inference

With the great success of convolutional neural networks (CNNs), interpretation of their internal network mechanism has been increasingly critical, while the network decision-making logic is still an open issue. In the bottom-up hierarchical logic of neuroscience, the decision-making process can be deduced from a series of sub-decision-making processes from low to high levels. Inspired by this, we propose the Concept-harmonized HierArchical INference (CHAIN) interpretation scheme. In CHAIN, a network decision-making process from shallow to deep layers is interpreted by the hierarchical backward inference based on visual concepts from high to low semantic levels. Firstly, we learned a general hierarchical visual-concept representation in CNN layered feature space by concept harmonizing model on a large concept dataset. Secondly, for interpreting a specific network decision-making process, we conduct the concept-harmonized hierarchical inference backward from the highest to the lowest semantic level. Specifically, the network learning for a target concept at a deeper layer is disassembled into that for concepts at shallower layers. Finally, a specific network decision-making process is explained as a form of concept-harmonized hierarchical inference, which is intuitively comparable to the bottom-up hierarchical visual recognition way. Quantitative and qualitative experiments demonstrate the effectiveness of the proposed CHAIN at both instance and class levels.

Brain Functional and Structural Alterations in Women With Bipolar Disorder and Suicidality

Objective: Suicide is the leading cause of death from bipolar disorder (BD). At least 25-50% of the patients with BD will attempt suicide, with suicide rates much higher in women patients than in men. It is crucial to explore the potential neural mechanism underlying suicidality in women with BD, which will lead to understanding and detection of suicidality and prevent death and injury from suicide. Methods: Brain function and structure were measured by amplitude of low-frequency fluctuation (ALFF) and gray matter volume (GMV) in 155 women [30 women with BD and a history of suicidality, 50 women with BD without suicidality, and 75 healthy controls (HC)]. The differences in ALFF and GMV across the BD with suicidality, BD without suicidality, and HC groups were investigated. Results: BD with suicidality showed significantly increased ALFF in the left and right cuneus compared with BD without suicidality and HC groups. Moreover, the GMV in the left lateral prefrontal cortex and left cuneus in BD with suicidality were significantly lower than those in BD without suicidality and HC groups, while the GMV of the right ventral prefrontal cortex was significantly decreased in both BD with and without suicidality groups. Conclusions: This study, combining functional and structural neuroimaging techniques, may help to identify specific pathophysiological changes in women with BD and suicidality. Increased ALFF and less GMV in cuneus might represent the neuroimaging features of suicidality in women with BD. Investigating this potential neuromarker for suicidality in women with BD may lead to the ability to prevent suicidality.

Depth in convolutional neural networks solves scene segmentation

Feed-forward deep convolutional neural networks (DCNNs) are, under specific conditions, matching and even surpassing human performance in object recognition in natural scenes. This performance suggests that the analysis of a loose collection of image features could support the recognition of natural object categories, without dedicated systems to solve specific visual subtasks. Research in humans however suggests that while feedforward activity may suffice for sparse scenes with isolated objects, additional visual operations ('routines') that aid the recognition process (e.g. segmentation or grouping) are needed for more complex scenes. Linking human visual processing to performance of DCNNs with increasing depth, we here explored if, how, and when object information is differentiated from the backgrounds they appear on. To this end, we controlled the information in both objects and backgrounds, as well as the relationship between them by adding noise, manipulating background congruence and systematically occluding parts of the image. Results indicate that with an increase in network depth, there is an increase in the distinction between object- and background information. For more shallow networks, results indicated a benefit of training on segmented objects. Overall, these results indicate that, de facto, scene segmentation can be performed by a network of sufficient depth. We conclude that the human brain could perform scene segmentation in the context of object identification without an explicit mechanism, by selecting or "binding" features that belong to the object and ignoring other features, in a manner similar to a very deep convolutional neural network.

Figure-Ground Organization in Natural Scenes: Performance of a Recurrent Neural Model Compared with Neurons of Area V2

A crucial step in understanding visual input is its organization into meaningful components, in particular object contours and partially occluded background structures. This requires that all contours are assigned to either the foreground or the background (border ownership assignment). While earlier studies showed that neurons in primate extrastriate cortex signal border ownership for simple geometric shapes, recent studies show consistent border ownership coding also for complex natural scenes. In order to understand how the brain performs this task, we developed a biologically plausible recurrent neural network that is fully image computable. Our model uses local edge detector ( B ) cells and grouping ( G ) cells whose activity represents proto-objects based on the integration of local feature information. G cells send modulatory feedback connections to those B cells that caused their activation, making the B cells border ownership selective. We found close agreement between our model and neurophysiological results in terms of the timing of border ownership signals (BOSs) as well as the consistency of BOSs across scenes. We also benchmarked our model on the Berkeley Segmentation Dataset and achieved performance comparable to recent state-of-the-art computer vision approaches. Our proposed model provides insight into the cortical mechanisms of figure-ground organization.

Getting "fumpered": Classifying objects by what has been done to them

Every object acquires its shape from some kind of generative process, such as manufacture, biological growth, or self-organization, in response to external forces. Inferring such generative processes from an observed shape is computationally challenging because a given process can lead to radically different shapes, and similar shapes can result from different generative processes. Here, we suggest that in some cases, generative processes endow objects with distinctive statistical features that observers can use to classify objects according to what has been done to them. We found that from the very first trials in an eight-alternative forced-choice classification task, observers were extremely good at classifying unfamiliar objects by the transformations that had shaped them. Further experiments show that the shape features underlying this ability are distinct from Euclidean shape similarity and that observers can separate and voluntarily respond to both aspects of objects. Our findings suggest that perceptual organization processes allow us to identify salient statistical shape features that are diagnostic of generative processes. By so doing, we can classify objects we have never seen before according to the processes that shaped them.

Recurrent Processing of Contour Integration in the Human Visual Cortex as Revealed By fMRI-Guided TMS

Contour integration is a critical step in visual perception because it groups discretely local elements into perceptually global contours. Previous investigations have suggested that striate and extrastriate visual areas are involved in this mid-level processing of visual perception. However, the temporal dynamics of these areas in the human brain during contour integration is less understood. The present study used functional magnetic resonance imaging-guided transcranial magnetic stimulation (TMS) to briefly disrupt 1 of 2 visual areas (V1/V2 and V3B) and examined the causal contributions of these areas to contour detection. The results demonstrated that the earliest critical time window at which behavioral detection performance was impaired by TMS pluses differed between V1/V2 and V3B. The first critical window of V3B (90-110 ms after stimulus onset) was earlier than that of V1/V2 (120-140 ms after stimulus onset), thus indicating that feedback connection from higher to lower area was necessary for complete contour integration. These results suggested that the fine processing of contour-related information in V1/V2 follows the generation of a coarse template in the higher visual areas, such as V3B. Our findings provide direct causal evidence that a recurrent mechanism is necessary for the integration of contours from cluttered background in the human brain.

Neural dynamics at successive stages of the ventral visual stream are consistent with hierarchical error signals

Ventral visual stream neural responses are dynamic, even for static image presentations. However, dynamical neural models of visual cortex are lacking as most progress has been made modeling static, time-averaged responses. Here, we studied population neural dynamics during face detection across three cortical processing stages. Remarkably,~30 milliseconds after the initially evoked response, we found that neurons in intermediate level areas decreased their responses to typical configurations of their preferred face parts relative to their response for atypical configurations even while neurons in higher areas achieved and maintained a preference for typical configurations. These hierarchical neural dynamics were inconsistent with standard feedforward circuits. Rather, recurrent models computing prediction errors between stages captured the observed temporal signatures. This model of neural dynamics, which simply augments the standard feedforward model of online vision, suggests that neural responses to static images may encode top-down prediction errors in addition to bottom-up feature estimates.

Image interpretation above and below the object level

Computational models of vision have advanced in recent years at a rapid rate, rivalling in some areas human-level performance. Much of the progress to date has focused on analysing the visual scene at the object level-the recognition and localization of objects in the scene. Human understanding of images reaches a richer and deeper image understanding both 'below' the object level, such as identifying and localizing object parts and sub-parts, as well as 'above' the object level, such as identifying object relations, and agents with their actions and interactions. In both cases, understanding depends on recovering meaningful structures in the image, and their components, properties and inter-relations, a process referred here as 'image interpretation'. In this paper, we describe recent directions, based on human and computer vision studies, towards human-like image interpretation, beyond the reach of current schemes, both below the object level, as well as some aspects of image interpretation at the level of meaningful configurations beyond the recognition of individual objects, and in particular, interactions between two people in close contact. In both cases the recognition process depends on the detailed interpretation of so-called 'minimal images', and at both levels recognition depends on combining 'bottom-up' processing, proceeding from low to higher levels of a processing hierarchy, together with 'top-down' processing, proceeding from high to lower levels stages of visual analysis.

A recurrent neural model for proto-object based contour integration and figure-ground segregation

Visual processing of objects makes use of both feedforward and feedback streams of information. However, the nature of feedback signals is largely unknown, as is the identity of the neuronal populations in lower visual areas that receive them. Here, we develop a recurrent neural model to address these questions in the context of contour integration and figure-ground segregation. A key feature of our model is the use of grouping neurons whose activity represents tentative objects ("proto-objects") based on the integration of local feature information. Grouping neurons receive input from an organized set of local feature neurons, and project modulatory feedback to those same neurons. Additionally, inhibition at both the local feature level and the object representation level biases the interpretation of the visual scene in agreement with principles from Gestalt psychology. Our model explains several sets of neurophysiological results (Zhou et al. Journal of Neuroscience, 20(17), 6594-6611 2000; Qiu et al. Nature Neuroscience, 10(11), 1492-1499 2007; Chen et al. Neuron, 82(3), 682-694 2014), and makes testable predictions about the influence of neuronal feedback and attentional selection on neural responses across different visual areas. Our model also provides a framework for understanding how object-based attention is able to select both objects and the features associated with them.

Perceptual integration rapidly activates dorsal visual pathway to guide local processing in early visual areas

Rapidly grouping local elements into an organized object (i.e., perceptual integration) is a fundamental yet challenging task, especially in noisy contexts. Previous studies demonstrate that ventral visual pathway, which is widely known to mediate object recognition, engages in the process by conveying object-level information processed in high-level areas to modulate low-level sensory areas. Meanwhile, recent evidence suggests that the dorsal visual pathway, which is not typically attributable to object recognition, is also involved in the process. However, the underlying whole-brain fine spatiotemporal neuronal dynamics remains unknown. Here we used magnetoencephalography (MEG) recordings in combination with a temporal response function (TRF) approach to dissociate the time-resolved neuronal response that specifically tracks the perceptual grouping course. We demonstrate that perceptual integration initiates robust and rapid responses along the dorsal visual pathway in a reversed hierarchical manner, faster than the ventral pathway. Specifically, the anterior intraparietal sulcus (IPS) responds first (i.e., within 100 ms), followed by activities backpropagating along the dorsal pathway to early visual areas (EVAs). The IPS activity causally modulates the EVA response, even when the global form information is task-irrelevant. The IPS-to-EVA response profile fails to appear when the global form could not be perceived. Our results support the crucial function of the dorsal visual pathway in perceptual integration, by quickly extracting a coarse global template (i.e., an initial object representation) within first 100 ms to guide subsequent local sensory processing so that the ambiguities in the visual inputs can be efficiently resolved.

How the brain integrates local elements into a global object (i.e., perceptual integration) in noisy contexts constitutes a fundamental yet challenging question in cognitive neuroscience. Here, we recorded brain activity by using magnetoencephalography from human subjects watching glass-pattern stimuli to examine the fine spatiotemporal neuronal responses during perceptual integration. We demonstrate that high-level brain regions initially extract a coarse global form of the inputs, which is then relayed along the dorsal visual pathway in a reversed hierarchical manner to low-level areas to modulate local analysis. This global-to-local modulation mechanism is especially beneficial in noisy environments by rapidly making an “initial guess” to guide detail analysis so that the ambiguities in inputs can be efficiently resolved.

Full interpretation of minimal images

The goal in this work is to model the process of 'full interpretation' of object images, which is the ability to identify and localize all semantic features and parts that are recognized by human observers. The task is approached by dividing the interpretation of the complete object to the interpretation of multiple reduced but interpretable local regions. In such reduced regions, interpretation is simpler, since the number of semantic components is small, and the variability of possible configurations is low. We model the interpretation process by identifying primitive components and relations that play a useful role in local interpretation by humans. To identify useful components and relations used in the interpretation process, we consider the interpretation of 'minimal configurations': these are reduced local regions, which are minimal in the sense that further reduction renders them unrecognizable and uninterpretable. We show that such minimal interpretable images have useful properties, which we use to identify informative features and relations used for full interpretation. We describe our interpretation model, and show results of detailed interpretations of minimal configurations, produced automatically by the model. Finally, we discuss possible extensions and implications of full interpretation to difficult visual tasks, such as recognizing social interactions, which are beyond the scope of current models of visual recognition.

Attention to distinguishing features in object recognition: An interactive-iterative framework

This article advances a framework that casts object recognition as a process of discrimination between alternative object identities, in which top-down and bottom-up processes interact-iteratively when necessary-with attention to distinguishing features playing a critical role. In two experiments, observers discriminated between different types of artificial fish. In parallel, a secondary, variable-SOA visual-probe detection task was used to examine the dynamics of visual attention. In Experiment 1, the fish varied in three distinguishing features: one indicating the general category (saltwater, freshwater), and one of the two other features indicating the specific type of fish within each category. As predicted, in the course of recognizing each fish, attention was allocated iteratively to the distinguishing features in an optimal manner: first to the general category feature, and then, based on its value, to the second feature that identified the specific fish. In Experiment 2, two types of fish could be discriminated on the basis of either of two distinguishing features, one more visually discriminable than the other. On some of the trials, one of the two alternative distinguishing features was occluded. As predicted, in the course of recognizing each fish, attention was directed initially to the more discriminable distinguishing feature, but when this feature was occluded, it was then redirected to the less discriminable feature. The implications of these findings, and the interactive-iterative framework they support, are discussed with regard to several fundamental issues having a long history in the literatures on object recognition, object categorization, and visual perception in general.

Object segmentation controls image reconstruction from natural scenes

The structure of the physical world projects images onto our eyes. However, those images are often poorly representative of environmental structure: well-defined boundaries within the eye may correspond to irrelevant features of the physical world, while critical features of the physical world may be nearly invisible at the retinal projection. The challenge for the visual cortex is to sort these two types of features according to their utility in ultimately reconstructing percepts and interpreting the constituents of the scene. We describe a novel paradigm that enabled us to selectively evaluate the relative role played by these two feature classes in signal reconstruction from corrupted images. Our measurements demonstrate that this process is quickly dominated by the inferred structure of the environment, and only minimally controlled by variations of raw image content. The inferential mechanism is spatially global and its impact on early visual cortex is fast. Furthermore, it retunes local visual processing for more efficient feature extraction without altering the intrinsic transduction noise. The basic properties of this process can be partially captured by a combination of small-scale circuit models and large-scale network architectures. Taken together, our results challenge compartmentalized notions of bottom-up/top-down perception and suggest instead that these two modes are best viewed as an integrated perceptual mechanism.

Biological vision is designed to discover the structure of the environment around us. To do this, it relies on ambiguous and often misleading information from the eyes: the boundary of a critical object may be invisible against a background of similar appearance, and may be overlooked in favour of the sharp contour projected by an irrelevant shadow. It remains unclear how human vision sorts different image features according to their relevance to the layout of objects within the scene. We demonstrate that vision achieves this goal via a specialized perceptual system for object segmentation that is one and the same with the feature extraction system: immediately after information is relayed to cortex by the eyes, the process of reconstructing image content from local features is controlled by a dedicated inferential mechanism that attempts to recover the underlying environmental structure; perception is quickly organized around the operation of this mechanism, which becomes the primary contextual influence on image reconstruction. The integrated nature of this perceptual mechanism defies current notions of separate top-down and bottom-up processes, offering a fresh view of how human vision operates on natural signals.

Interactions between feedback and lateral connections in the primary visual cortex

Perceptual grouping of line segments into object contours has been thought to be mediated, in part, by long-range horizontal connectivity intrinsic to the primary visual cortex (V1), with a contribution by top-down feedback projections. To dissect the contributions of intraareal and interareal connections during contour integration, we applied conditional Granger causality analysis to assess directional influences among neural signals simultaneously recorded from visual cortical areas V1 and V4 of monkeys performing a contour detection task. Our results showed that discounting the influences from V4 markedly reduced V1 lateral interactions, indicating dependence on feedback signals of the effective connectivity within V1. On the other hand, the feedback influences were reciprocally dependent on V1 lateral interactions because the modulation strengths from V4 to V1 were greatly reduced after discounting the influences from other V1 neurons. Our findings suggest that feedback and lateral connections closely interact to mediate image grouping and segmentation.

Responses in early visual areas to contour integration are context dependent

It has been shown that early visual areas are involved in contour processing. However, it is not clear how local and global context interact to influence responses in those areas, nor has the interarea coordination that yields coherent structural percepts been fully studied, especially in human observers. In this study, we used functional magnetic resonance imaging (fMRI) to measure activity in early visual cortex while observers performed a contour detection task in which alignment of Gabor elements and background clutter were manipulated. Six regions of interest (two regions, containing either the cortex representing the target or the background clutter, in each of areas V1, V2, and V3) were predefined using separate target versus background functional localizer scans. The first analysis using a general linear model showed that in the presence of background clutter, responses in V1 and V2 target regions of interest were significantly stronger to aligned than unaligned contours, whereas when background clutter was absent, no significant difference was observed. The second analysis using interarea correlations showed that with background clutter, there was an increase in V1–V2 coordination within the target regions when perceiving aligned versus unaligned contours; without clutter, however, correlations between V1 and V2 were similar no matter whether aligned contours were present or not. Both the average response magnitude and the connectivity analysis suggest different mechanisms support contour processing with or without background distractors. Coordination between V1 and V2 may play a major role in coherent structure perception, especially with complex scene organization.

Serial grouping of 2D-image regions with object-based attention in humans

After an initial stage of local analysis within the retina and early visual pathways, the human visual system creates a structured representation of the visual scene by co-selecting image elements that are part of behaviorally relevant objects. The mechanisms underlying this perceptual organization process are only partially understood. We here investigate the time-course of perceptual grouping of two-dimensional image-regions by measuring the reaction times of human participants and report that it is associated with the gradual spread of object-based attention. Attention spreads fastest over large and homogeneous areas and is slowed down at locations that require small-scale processing. We find that the time-course of the object-based selection process is well explained by a 'growth-cone' model, which selects surface elements in an incremental, scale-dependent manner. We discuss how the visual cortical hierarchy can implement this scale-dependent spread of object-based attention, leveraging the different receptive field sizes in distinct cortical areas.

DOI:http://dx.doi.org/10.7554/eLife.14320.001

When we look at an object, we perceive it as a whole. However, this is not how the brain processes objects. Instead, cells at early stages of the visual system respond selectively to single features of the object, such as edges. Moreover, each cell responds to its target feature in only a small region of space known as its receptive field. At higher levels of the visual system, cells respond to more complex features: angles rather than edges, for example. The receptive fields of the cells are also larger. For us to see an object, the brain must therefore 'stitch' together diverse features into a unified impression.

This process is termed perceptual grouping. But how does it work? Jeurissen et al. hypothesized that this process depends on the visual system’s attention spreading over a region in the image occupied by an object, and that the speed of the process will depend on the size of the receptive fields involved. If an image region is narrow, the visual system must recruit cells with small receptive fields to process the individual features. Grouping will therefore be slow. By contrast, if the object consists of large uniform areas lacking in detail, grouping should be fast. These assumptions give rise to a model called the “growth-conemodel”, which makes a number of specific predictions about reaction times during perceptual grouping.

Jeurissen et al. tested the growth-cone model’s predictions by measuring the speed of perceptual grouping in 160 human volunteers. These volunteers looked at an image made up of two simple shapes, and reported whether two dots fell on the same or different shapes. The results supported the growth-cone model. People were able to group large and uniform areas quickly, but were slower for narrow areas. Grouping also took more time when the distance between the dots increased. Hence, perceptual grouping of everyday objects calls on a step-by-step process that resembles solving a small maze.

The results also revealed that perceptual grouping of simple shapes relies on the spreading of visual attention over the relevant object. Furthermore, the data support the hypothesis that perceptual grouping makes use of the different sizes of receptive fields at various levels of the visual system. Further research will be needed to translate these findings to the more complex natural scenes we encounter in our daily lives.

DOI:http://dx.doi.org/10.7554/eLife.14320.002

What insights can fMRI offer into the structure and function of mid-tier visual areas?

Inferring neural responses from functional magnetic resonance imaging (fMRI) data is challenging. Even if we take advantage of high-field systems to acquire data with submillimeter resolution, we are still acquiring data in which a single datum summarizes the responses of tens of thousands of neurons. Excitation and inhibition, spikes and subthreshold membrane potential modulations, local and long-range computations, and tuned and nonselective responses are mixed together in one signal. With a priori knowledge of the underlying neural population responses, careful experiment design allows us to manipulate the experiment or task design so that subpopulations are selectively modulated, and our experiments can reveal those tuning functions. However, because we want to be able to use fMRI to discover new kinds of tuning functions and selectivity, we cannot limit ourselves to experiments in which we already know what we are looking for. Broadly speaking, analyses that rely on classification of responses that are distributed across the local neural population [multi-voxel pattern analyses (MVPA)] offer the ability to discover new kinds of information representation and selectivities in neural subpopulations. There is, however, no way to determine how the information discovered with MVPA or other analyses is related to the underlying neuronal tuning functions. Therefore, we must continue to rely on behavioral, computational, and animal models to develop theories of information representation in mid-tier visual cortical areas. Once encoding models exist, fMRI can be powerful for testing these a priori models of information representation. As an aide in developing these models, an important contribution that fMRI can make to our understanding of mid-tier visual areas is derived from connectivity analyses and experiments that study information sharing between visual areas. This ability to quantify localized population average responses throughout the brain is the strength we can best leverage to discover new properties of local and long-range neural networks.

Electrophysiological correlates of top-down effects facilitating natural image categorization are disrupted by the attenuation of low spatial frequency information

The modulatory effects of low and high spatial frequencies on the posterior C1, P1 and N1 event-related potential (ERP) amplitudes have long been known from previous electrophysiological studies. There is also evidence that categorization of complex natural images relies on top-down processes, probably by facilitating contextual associations during the recognition process. However, to our knowledge, no study has investigated so far how such top-down effects are manifested in scalp ERPs, when presenting natural images with attenuated low or high spatial frequency information. Twenty-one healthy subjects participated in an animal vs. vehicle categorization task with intact grayscale stimuli and images predominantly containing high (HSF) or low spatial frequencies (LSF). ERP scalp maps and amplitudes/latencies measured above occipital, parietal and frontocentral sites were compared among the three stimulus conditions. Although early occipital components (C1 and P1) were modulated by spatial frequencies, the time range of the N1 was the earliest to show top-down effects for images with unmodified low spatial frequency spectrum (intact and LSF stimuli). This manifested in ERP amplitude changes spreading to anterior scalp sites and shorter posterior N1 latencies. Finally, the frontocentral N350 and the centroparietal LPC were differently influenced by spatial frequency filtering, with the LPC being the only component to show an amplitude and latency modulation congruent with the behavioral responses (sensitivity index and reaction times). Our results strengthen the coarse-to-fine model of object recognition and provide electrophysiological evidence for low spatial frequency-based top-down effects within the first 200 ms of visual processing.

Increased alpha band activity indexes inhibitory competition across a border during figure assignment

Figure-ground assignment is thought to entail inhibitory competition between potential objects on opposite sides of a shared border; the winner is perceived as the figure, and the loser as the shapeless ground. Computational models and response time measures support this understanding but to date no online measure of inhibitory competition during figure-ground assignment has been reported. The current study assays electroencephalogram (EEG) alpha power as a measure of inhibitory competition during figure-ground assignment. Activity in the EEG alpha band has been linked to functional inhibition in the brain, and it has been proposed that increased alpha power reflects increased inhibition. In 2 experiments participants viewed silhouettes designed so that the insides would be perceived as figures. Real-world silhouettes depicted namable objects. Novel silhouettes depicted novel objects on the insides of their borders, but varied in the amount of hypothesized cross-border competition for figural status: In "Low-Competition" silhouettes, the borders suggested novel objects on the outside as well as on the inside. In "High-Competition" silhouettes the borders suggested portions of real-world objects on the outside; these compete with the figural properties favoring the inside as figure. Participants accurately categorized both types of novel silhouettes as "novel" objects and were unaware of the real world objects suggested on the outside of the High-Competition silhouettes. In both experiments, we observed more alpha power while participants viewed High- rather than Low-Competition novel silhouettes. These are the first results to show via an online index of neural activity that figure assignment entails inhibitory competition.

On the role of spatial phase and phase correlation in vision, illusion, and cognition

Numerous findings indicate that spatial phase bears an important cognitive information. Distortion of phase affects topology of edge structures and makes images unrecognizable. In turn, appropriately phase-structured patterns give rise to various illusions of virtual image content and apparent motion. Despite a large body of phenomenological evidence not much is known yet about the role of phase information in neural mechanisms of visual perception and cognition. Here, we are concerned with analysis of the role of spatial phase in computational and biological vision, emergence of visual illusions and pattern recognition. We hypothesize that fundamental importance of phase information for invariant retrieval of structural image features and motion detection promoted development of phase-based mechanisms of neural image processing in course of evolution of biological vision. Using an extension of Fourier phase correlation technique, we show that the core functions of visual system such as motion detection and pattern recognition can be facilitated by the same basic mechanism. Our analysis suggests that emergence of visual illusions can be attributed to presence of coherently phase-shifted repetitive patterns as well as the effects of acuity compensation by saccadic eye movements. We speculate that biological vision relies on perceptual mechanisms effectively similar to phase correlation, and predict neural features of visual pattern (dis)similarity that can be used for experimental validation of our hypothesis of “cognition by phase correlation.”

The functional architecture of the ventral temporal cortex and its role in categorization

Visual categorization is thought to occur in the human ventral temporal cortex (VTC), but how this categorization is achieved is still largely unknown. In this Review, we consider the computations and representations that are necessary for categorization and examine how the microanatomical and macroanatomical layout of the VTC might optimize them to achieve rapid and flexible visual categorization. We propose that efficient categorization is achieved by organizing representations in a nested spatial hierarchy in the VTC. This spatial hierarchy serves as a neural infrastructure for the representational hierarchy of visual information in the VTC and thereby enables flexible access to category information at several levels of abstraction.

Incremental integration of global contours through interplay between visual cortical areas

The traditional view on visual processing emphasizes a hierarchy: local line segments are first linked into global contours, which in turn are assembled into more complex forms. Distinct from this bottom-up viewpoint, here we provide evidence for a theoretical framework whereby objects and their parts are processed almost concurrently in a bidirectional cortico-cortical loop. By simultaneous recordings from V1 and V4 in awake monkeys, we found that information about global contours in a cluttered background emerged initially in V4, started ∼40 ms later in V1, and continued to develop in parallel in both areas. Detailed analysis of neuronal response properties implicated contour integration to emerge from both bottom-up and reentrant processes. Our results point to an incremental integration mechanism: feedforward assembling accompanied by feedback disambiguating to define and enhance the global contours and to suppress background noise. The consequence is a parallel accumulation of contour information over multiple cortical areas.

Regions of mid-level human visual cortex sensitive to the global coherence of local image patches

The global structural arrangement and spatial layout of the visual environment must be derived from the integration of local signals represented in the lower tiers of the visual system. This interaction between the spatially local and global properties of visual stimulation underlies many of our visual capacities, and how this is achieved in the brain is a central question for visual and cognitive neuroscience. Here, we examine the sensitivity of regions of the posterior human brain to the global coordination of spatially displaced naturalistic image patches. We presented observers with image patches in two circular apertures to the left and right of central fixation, with the patches drawn from either the same (coherent condition) or different (noncoherent condition) extended image. Using fMRI at 7T (n = 5), we find that global coherence affected signal amplitude in regions of dorsal mid-level cortex. Furthermore, we find that extensive regions of mid-level visual cortex contained information in their local activity pattern that could discriminate coherent and noncoherent stimuli. These findings indicate that the global coordination of local naturalistic image information has important consequences for the processing in human mid-level visual cortex.

Interaction between dorsal and ventral processing streams: where, when and how?

The execution of complex visual, auditory, and linguistic behaviors requires a dynamic interplay between spatial ('where/how') and non-spatial ('what') information processed along the dorsal and ventral processing streams. However, while it is acknowledged that there must be some degree of interaction between the two processing networks, how they interact, both anatomically and functionally, is a question which remains little explored. The current review examines the anatomical, temporal, and behavioral evidence regarding three potential models of dual stream interaction: (1) computations along the two pathways proceed independently and in parallel, reintegrating within shared target brain regions; (2) processing along the separate pathways is modulated by the existence of recurrent feedback loops; and (3) information is transferred directly between the two pathways at multiple stages and locations along their trajectories.

Vision: are models of object recognition catching up with the brain?

Object recognition has been a central yet elusive goal of computational vision. For many years, computer performance seemed highly deficient and unable to emulate the basic capabilities of the human recognition system. Over the past decade or so, computer scientists and neuroscientists have developed algorithms and systems-and models of visual cortex-that have come much closer to human performance in visual identification and categorization. In this personal perspective, we discuss the ongoing struggle of visual models to catch up with the visual cortex, identify key reasons for the relatively rapid improvement of artificial systems and models, and identify open problems for computational vision in this domain.

Compressive spatial summation in human visual cortex

Neurons within a small (a few cubic millimeters) region of visual cortex respond to stimuli within a restricted region of the visual field. Previous studies have characterized the population response of such neurons using a model that sums contrast linearly across the visual field. In this study, we tested linear spatial summation of population responses using blood oxygenation level-dependent (BOLD) functional MRI. We measured BOLD responses to a systematic set of contrast patterns and discovered systematic deviation from linearity: the data are more accurately explained by a model in which a compressive static nonlinearity is applied after linear spatial summation. We found that the nonlinearity is present in early visual areas (e.g., V1, V2) and grows more pronounced in relatively anterior extrastriate areas (e.g., LO-2, VO-2). We then analyzed the effect of compressive spatial summation in terms of changes in the position and size of a viewed object. Compressive spatial summation is consistent with tolerance to changes in position and size, an important characteristic of object representation.

Consequences of polar form coherence for fMRI responses in human visual cortex

Relevant features in the visual image are often spatially extensive and have complex orientation structure. Our perceptual sensitivity to such spatial form is demonstrated by polar Glass patterns, in which an array of randomly-positioned dot pairs that are each aligned with a particular polar displacement (rotation, for example) yield a salient impression of spatial structure. Such patterns are typically considered to be processed in two main stages: local spatial filtering in low-level visual cortex followed by spatial pooling and complex form selectivity in mid-level visual cortex. However, it remains unclear both whether reciprocal interactions within the cortical hierarchy are involved in polar Glass pattern processing and which mid-level areas identify and communicate polar Glass pattern structure. Here, we used functional magnetic resonance imaging (fMRI) at 7T to infer the magnitude of neural response within human low-level and mid-level visual cortex to polar Glass patterns of varying coherence (proportion of signal elements). The activity within low-level visual areas V1 and V2 was not significantly modulated by polar Glass pattern coherence, while the low-level area V3, dorsal and ventral mid-level areas, and the human MT complex each showed a positive linear coherence response functions. The cortical processing of polar Glass patterns thus appears to involve primarily feedforward communication of local signals from V1 and V2, with initial polar form selectivity reached in V3 and distributed to multiple pathways in mid-level visual cortex.

Top-down feedback in an HMAX-like cortical model of object perception based on hierarchical Bayesian networks and belief propagation

Hierarchical generative models, such as Bayesian networks, and belief propagation have been shown to provide a theoretical framework that can account for perceptual processes, including feedforward recognition and feedback modulation. The framework explains both psychophysical and physiological experimental data and maps well onto the hierarchical distributed cortical anatomy. However, the complexity required to model cortical processes makes inference, even using approximate methods, very computationally expensive. Thus, existing object perception models based on this approach are typically limited to tree-structured networks with no loops, use small toy examples or fail to account for certain perceptual aspects such as invariance to transformations or feedback reconstruction. In this study we develop a Bayesian network with an architecture similar to that of HMAX, a biologically-inspired hierarchical model of object recognition, and use loopy belief propagation to approximate the model operations (selectivity and invariance). Crucially, the resulting Bayesian network extends the functionality of HMAX by including top-down recursive feedback. Thus, the proposed model not only achieves successful feedforward recognition invariant to noise, occlusions, and changes in position and size, but is also able to reproduce modulatory effects such as illusory contour completion and attention. Our novel and rigorous methodology covers key aspects such as learning using a layerwise greedy algorithm, combining feedback information from multiple parents and reducing the number of operations required. Overall, this work extends an established model of object recognition to include high-level feedback modulation, based on state-of-the-art probabilistic approaches. The methodology employed, consistent with evidence from the visual cortex, can be potentially generalized to build models of hierarchical perceptual organization that include top-down and bottom-up interactions, for example, in other sensory modalities.

Design Democratization with Communities

The authors present community drawing as meaningful representations to inform locally valid technology design. They investigate recognition within and across cultural borders, thereby exposing variances of localities. The study contributes to the still scarce body of empirical work on culturally meaningful development of visual representations and recognition, as part of a longitudinal research project in which we co-design a 3D visualization for a specific Namibian pilot site.

Predictive top-down integration of prior knowledge during speech perception

A striking feature of human perception is that our subjective experience depends not only on sensory information from the environment but also on our prior knowledge or expectations. The precise mechanisms by which sensory information and prior knowledge are integrated remain unclear, with longstanding disagreement concerning whether integration is strictly feedforward or whether higher-level knowledge influences sensory processing through feedback connections. Here we used concurrent EEG and MEG recordings to determine how sensory information and prior knowledge are integrated in the brain during speech perception. We manipulated listeners' prior knowledge of speech content by presenting matching, mismatching, or neutral written text before a degraded (noise-vocoded) spoken word. When speech conformed to prior knowledge, subjective perceptual clarity was enhanced. This enhancement in clarity was associated with a spatiotemporal profile of brain activity uniquely consistent with a feedback process: activity in the inferior frontal gyrus was modulated by prior knowledge before activity in lower-level sensory regions of the superior temporal gyrus. In parallel, we parametrically varied the level of speech degradation, and therefore the amount of sensory detail, so that changes in neural responses attributable to sensory information and prior knowledge could be directly compared. Although sensory detail and prior knowledge both enhanced speech clarity, they had an opposite influence on the evoked response in the superior temporal gyrus. We argue that these data are best explained within the framework of predictive coding in which sensory activity is compared with top-down predictions and only unexplained activity propagated through the cortical hierarchy.

Spatially invariant computations in stereoscopic vision

PERCEPTION OF STEREOSCOPIC DEPTH REQUIRES THAT VISUAL SYSTEMS SOLVE A CORRESPONDENCE PROBLEM: find parts of the left-eye view of the visual scene that correspond to parts of the right-eye view. The standard model of binocular matching implies that similarity of left and right images is computed by inter-ocular correlation. But the left and right images of the same object are normally distorted relative to one another by the binocular projection, in particular when slanted surfaces are viewed from close distance. Correlation often fails to detect correct correspondences between such image parts. We investigate a measure of inter-ocular similarity that takes advantage of spatially invariant computations similar to the computations performed by complex cells in biological visual systems. This measure tolerates distortions of corresponding image parts and yields excellent performance over a much larger range of surface slants than the standard model. The results suggest that, rather than serving as disparity detectors, multiple binocular complex cells take part in the computation of inter-ocular similarity, and that visual systems are likely to postpone commitment to particular binocular disparities until later stages in the visual process.

Object recognition in clutter: cortical responses depend on the type of learning

Theoretical studies suggest that the visual system uses prior knowledge of visual objects to recognize them in visual clutter, and posit that the strategies for recognizing objects in clutter may differ depending on whether or not the object was learned in clutter to begin with. We tested this hypothesis using functional magnetic resonance imaging (fMRI) of human subjects. We trained subjects to recognize naturalistic, yet novel objects in strong or weak clutter. We then tested subjects' recognition performance for both sets of objects in strong clutter. We found many brain regions that were differentially responsive to objects during object recognition depending on whether they were learned in strong or weak clutter. In particular, the responses of the left fusiform gyrus (FG) reliably reflected, on a trial-to-trial basis, subjects' object recognition performance for objects learned in the presence of strong clutter. These results indicate that the visual system does not use a single, general-purpose mechanism to cope with clutter. Instead, there are two distinct spatial patterns of activation whose responses are attributable not to the visual context in which the objects were seen, but to the context in which the objects were learned.

Rapid Interactions between Lexical Semantic and Word Form Analysis during Word Recognition in Context: Evidence from ERPs

We used ERPs to investigate the time course of interactions between lexical semantic and sublexical visual word form processing during word recognition. Participants read sentence-embedded pseudowords that orthographically resembled a contextually supported real word (e.g., “She measured the flour so she could bake a ceke…”) or did not (e.g., “She measured the flour so she could bake a tont…”) along with nonword consonant strings (e.g., “She measured the flour so she could bake a srdt…”). Pseudowords that resembled a contextually supported real word (“ceke”) elicited an enhanced positivity at 130 msec (P130), relative to real words (e.g., “She measured the flour so she could bake a cake…”). Pseudowords that did not resemble a plausible real word (“tont”) enhanced the N170 component, as did nonword consonant strings (“srdt”). The effect pattern shows that the visual word recognition system is, perhaps, counterintuitively, more rapidly sensitive to minor than to flagrant deviations from contextually predicted inputs. The findings are consistent with rapid interactions between lexical and sublexical representations during word recognition, in which rapid lexical access of a contextually supported word (CAKE) provides top–down excitation of form features (“cake”), highlighting the anomaly of an unexpected word “ceke.”

What are the Visual Features Underlying Rapid Object Recognition?

Research progress in machine vision has been very significant in recent years. Robust face detection and identification algorithms are already readily available to consumers, and modern computer vision algorithms for generic object recognition are now coping with the richness and complexity of natural visual scenes. Unlike early vision models of object recognition that emphasized the role of figure-ground segmentation and spatial information between parts, recent successful approaches are based on the computation of loose collections of image features without prior segmentation or any explicit encoding of spatial relations. While these models remain simplistic models of visual processing, they suggest that, in principle, bottom-up activation of a loose collection of image features could support the rapid recognition of natural object categories and provide an initial coarse visual representation before more complex visual routines and attentional mechanisms take place. Focusing on biologically plausible computational models of (bottom-up) pre-attentive visual recognition, we review some of the key visual features that have been described in the literature. We discuss the consistency of these feature-based representations with classical theories from visual psychology and test their ability to account for human performance on a rapid object categorization task.

Adaptive shape processing in primary visual cortex

The ability to derive meaning from complex sensory input requires the integration of information over space and time, as well as cognitive mechanisms to shape that integration. We studied these processes in the primary visual cortex (V1), where neurons are thought to integrate visual inputs along contours defined by an association field (AF). We recorded extracellularly from single cells in macaque V1 to map the AF, by using an optimization algorithm to find the contours that maximally activated individual cells. We combined the algorithm with a delayed-match-to-sample task, to test how the optimal contours might be molded by the monkey's expectation for particular cue shapes. We found that V1 neurons were selective for complex shapes, a property previously ascribed to higher cortical areas. Furthermore, the shape selectivity was reprogrammed by perceptual task: Over the whole network, the optimal modes of geometric selectivity shifted between distinct subsets of the AF, alternately representing different stimulus features known to predominate in natural scenes. Our results suggest a general model of cortical function, whereby horizontal connections provide a broad domain of potential associations, and top-down inputs dynamically gate these linkages to task switch the function of a network.

On the emergence of primary visual perception

We propose a concise novel conceptual and biological framework for the analysis of primary visual perception (PVP) that refers to the most basic levels of our awake subjective visual experiences. Neural representations for image content elaborated within V1/V2 and the early occipitotemporal (ventral) loop remain only latent with respect to PVP until spatially localized with respect to an attending observer. This process requires more than the downstream deployment of attentional resources onto targeted neurons. Additionally, the source neurons for such processes must be linked to a neural representation subserving a first-person perspective. We hypothesize that the simultaneous emergence of both the perceptual experience of image content and the personal inference of its ownership requires the resolution of any conflicting neuronal signaling between afferent and recurrent projections within and between both the ventral and dorsal streams. The V1/V2 complex and ventral cortical areas V3 and the V4 complex together with dorsal cortical areas LIP, VIP, and 7a with additional contributions from the motion areas V5/MT (middle temporal area), FST (fundus of superior temporal area), and MST (medial superior temporal area) together with their subcortical dependencies have the physiological properties required to constitute a "posterior perceptual core" that encodes the normal primary perceptual experience of image content, space, and sense of minimal self.

A link between visual disambiguation and visual memory

Sensory information in the retinal image is typically too ambiguous to support visual object recognition by itself. Theories of visual disambiguation posit that to disambiguate, and thus interpret, the incoming images, the visual system must integrate the sensory information with previous knowledge of the visual world. However, the underlying neural mechanisms remain unclear. Using functional magnetic resonance imaging (fMRI) of human subjects, we have found evidence for functional specialization for storing disambiguating information in memory versus interpreting incoming ambiguous images. Subjects viewed two-tone, "Mooney" images, which are typically ambiguous when seen for the first time but are quickly disambiguated after viewing the corresponding unambiguous color images. Activity in one set of regions, including a region in the medial parietal cortex previously reported to play a key role in Mooney image disambiguation, closely reflected memory for previously seen color images but not the subsequent disambiguation of Mooney images. A second set of regions, including the superior temporal sulcus, showed the opposite pattern, in that their responses closely reflected the subjects' percepts of the disambiguated Mooney images on a stimulus-to-stimulus basis but not the memory of the corresponding color images. Functional connectivity between the two sets of regions was stronger during those trials in which the disambiguated percept was stronger. This functional interaction between brain regions that specialize in storing disambiguating information in memory versus interpreting incoming ambiguous images may represent a general mechanism by which previous knowledge disambiguates visual sensory information.

What and where: a Bayesian inference theory of attention

In the theoretical framework of this paper, attention is part of the inference process that solves the visual recognition problem of what is where. The theory proposes a computational role for attention and leads to a model that predicts some of its main properties at the level of psychophysics and physiology. In our approach, the main goal of the visual system is to infer the identity and the position of objects in visual scenes: spatial attention emerges as a strategy to reduce the uncertainty in shape information while feature-based attention reduces the uncertainty in spatial information. Featural and spatial attention represent two distinct modes of a computational process solving the problem of recognizing and localizing objects, especially in difficult recognition tasks such as in cluttered natural scenes. We describe a specific computational model and relate it to the known functional anatomy of attention. We show that several well-known attentional phenomena--including bottom-up pop-out effects, multiplicative modulation of neuronal tuning curves and shift in contrast responses--all emerge naturally as predictions of the model. We also show that the Bayesian model predicts well human eye fixations (considered as a proxy for shifts of attention) in natural scenes.

Task-specific codes for face recognition: how they shape the neural representation of features for detection and individuation

BACKGROUND

The variety of ways in which faces are categorized makes face recognition challenging for both synthetic and biological vision systems. Here we focus on two face processing tasks, detection and individuation, and explore whether differences in task demands lead to differences both in the features most effective for automatic recognition and in the featural codes recruited by neural processing.

METHODOLOGY/PRINCIPAL FINDINGS

Our study appeals to a computational framework characterizing the features representing object categories as sets of overlapping image fragments. Within this framework, we assess the extent to which task-relevant information differs across image fragments. Based on objective differences we find among task-specific representations, we test the sensitivity of the human visual system to these different face descriptions independently of one another. Both behavior and functional magnetic resonance imaging reveal effects elicited by objective task-specific levels of information. Behaviorally, recognition performance with image fragments improves with increasing task-specific information carried by different face fragments. Neurally, this sensitivity to the two tasks manifests as differential localization of neural responses across the ventral visual pathway. Fragments diagnostic for detection evoke larger neural responses than non-diagnostic ones in the right posterior fusiform gyrus and bilaterally in the inferior occipital gyrus. In contrast, fragments diagnostic for individuation evoke larger responses than non-diagnostic ones in the anterior inferior temporal gyrus. Finally, for individuation only, pattern analysis reveals sensitivity to task-specific information within the right "fusiform face area".

CONCLUSIONS/SIGNIFICANCE

OUR RESULTS DEMONSTRATE: 1) information diagnostic for face detection and individuation is roughly separable; 2) the human visual system is independently sensitive to both types of information; 3) neural responses differ according to the type of task-relevant information considered. More generally, these findings provide evidence for the computational utility and the neural validity of fragment-based visual representation and recognition.