Iris I. A. Groen

Making Sense of Real-World Scenes

To interact with the world, we have to make sense of the continuous sensory input conveying information about our environment. A recent surge of studies has investigated the processes enabling scene understanding, using increasingly complex stimuli and sophisticated analyses to highlight the visual features and brain regions involved. However, there are two major challenges to producing a comprehensive framework for scene understanding. First, scene perception is highly dynamic, subserving multiple behavioral goals. Second, a multitude of different visual properties co-occur across scenes and may be correlated or independent. We synthesize the recent literature and argue that for a complete view of scene understanding, it is necessary to account for both differing observer goals and the contribution of diverse scene properties.

Contributions of low- and high-level properties to neural processing of visual scenes in the human brain

Visual scene analysis in humans has been characterized by the presence of regions in extrastriate cortex that are selectively responsive to scenes compared with objects or faces. While these regions have often been interpreted as representing high-level properties of scenes (e.g. category), they also exhibit substantial sensitivity to low-level (e.g. spatial frequency) and mid-level (e.g. spatial layout) properties, and it is unclear how these disparate findings can be united in a single framework. In this opinion piece, we suggest that this problem can be resolved by questioning the utility of the classical low- to high-level framework of visual perception for scene processing, and discuss why low- and mid-level properties may be particularly diagnostic for the behavioural goals specific to scene perception as compared to object recognition. In particular, we highlight the contributions of low-level vision to scene representation by reviewing (i) retinotopic biases and receptive field properties of scene-selective regions and (ii) the temporal dynamics of scene perception that demonstrate overlap of low- and mid-level feature representations with those of scene category. We discuss the relevance of these findings for scene perception and suggest a more expansive framework for visual scene analysis. This article is part of the themed issue ‘Auditory and visual scene analysis’.

Evaluating the correspondence between face-, scene-, and object-selectivity and retinotopic organization within lateral occipitotemporal cortex.

The organization of human lateral occipitotemporal cortex (lOTC) has been characterized largely according to two distinct principles: retinotopy and category-selectivity. Whereas category-selective regions were originally thought to exist beyond retinotopic maps, recent evidence highlights overlap. Here, we combined detailed mapping of retinotopy, using population receptive fields (pRF), and category-selectivity to examine and contrast the retinotopic profiles of scene- (occipital place area, OPA), face- (occipital face area, OFA) and object- (lateral occipital cortex, LO) selective regions of lOTC. We observe striking differences in the relationship each region has to underlying retinotopy. Whereas OPA overlapped multiple retinotopic maps (including V3A, V3B, LO1, and LO2), and LO overlapped two maps (LO1 and LO2), OFA overlapped almost none. There appears no simple consistent relationship between category-selectivity and retinotopic maps, meaning category-selective regions are not constrained spatially to retinotopic map borders consistently. The multiple maps that overlap OPA suggests it is likely not appropriate to conceptualize it as a single scene-selective region, whereas the inconsistency in any systematic map overlapping OFA suggests it may constitute a more uniform area. Beyond their relationship to retinotopy, all three regions evidenced strongly retinotopic voxels, with pRFs exhibiting a significant bias towards the contralateral lower visual field, despite differences in pRF size, contributing to an emerging literature suggesting this bias is present across much of lOTC. Taken together, these results suggest that whereas category-selective regions are not constrained to consistently contain ordered retinotopic maps, they nonetheless likely inherit retinotopic characteristics of the maps from which they draw information.

Distinct contributions of functional and deep neural network features to representational similarity of scenes in human brain and behavior

Inherent correlations between visual and semantic features in real-world scenes make it difficult to determine how different scene properties contribute to neural representations. Here, we assessed the contributions of multiple properties to scene representation by partitioning the variance explained in human behavioral and brain measurements by three feature models whose inter-correlations were minimized a priori through stimulus preselection. Behavioral assessments of scene similarity reflected unique contributions from a functional feature model indicating potential actions in scenes as well as high-level visual features from a deep neural network (DNN). In contrast, similarity of cortical responses in scene-selective areas was uniquely explained by mid- and high-level DNN features only, while an object label model did not contribute uniquely to either domain. The striking dissociation between functional and DNN features in their contribution to behavioral and brain representations of scenes indicates that scene-selective cortex represents only a subset of behaviorally relevant scene information.

The temporal evolution of conceptual object representations revealed through models of behavior, semantics and deep neural networks

Abstract Visual object representations are commonly thought to emerge rapidly, yet it has remained unclear to what extent early brain responses reflect purely low‐level visual features of these objects and how strongly those features contribute to later categorical or conceptual representations. Here, we aimed to estimate a lower temporal bound for the emergence of conceptual representations by defining two criteria that characterize such representations: 1) conceptual object representations should generalize across different exemplars of the same object, and 2) these representations should reflect high‐level behavioral judgments. To test these criteria, we compared magnetoencephalography (MEG) recordings between two groups of participants (n = 16 per group) exposed to different exemplar images of the same object concepts. Further, we disentangled low‐level from high‐level MEG responses by estimating the unique and shared contribution of models of behavioral judgments, semantics, and different layers of deep neural networks of visual object processing. We find that 1) both generalization across exemplars as well as generalization of object‐related signals across time increase after 150 ms, peaking around 230 ms; 2) representations specific to behavioral judgments emerged rapidly, peaking around 160 ms. Collectively, these results suggest a lower bound for the emergence of conceptual object representations around 150 ms following stimulus onset. HighlightsUsed MEG to reveal lower bound for emergence of conceptual object representations.Two criteria: between‐exemplar generalization and relationship to behavior.MEG pattern similarity between exemplars rises after 160 ms.Model of behavior explains unique MEG variance after 150 ms.Earlier MEG response well captured by early layer of deep neural network model.

Scene complexity modulates degree of feedback activity during object recognition in natural scenes

Object recognition is thought to be mediated by rapid feed-forward activation of object-selective cortex, with limited contribution of feedback. However, disruption of visual evoked activity beyond feed-forward processing stages has been demonstrated to affect object recognition performance. Here, we unite these findings by reporting that the detection of target objects in natural scenes is selectively characterized by enhanced feedback when these objects are embedded in high complexity scenes. Human participants performed an animal target detection task on scenes with low, medium or high complexity as determined by a biologically plausible computational model of low-level contrast statistics. Three converging lines of evidence indicate that feedback was enhanced during categorization of scenes with high, but not low or medium complexity. First, functional magnetic resonance imaging (fMRI) activity in early visual cortex (V1) was selectively enhanced for target objects in scenes with high complexity. Second, event-related potentials (ERPs) evoked by high complexity scenes were selectively enhanced from 220 ms after stimulus-onset. Third, behavioral performance deteriorated for highly complex scenes when participants were pressed for time, but not when they could process the scenes fully and thereby benefit from the enhanced feedback. Formal modeling of the reaction time distributions revealed that object information accumulated more slowly for high complexity scenes (resulting in more errors especially for fast decisions), and directly related to the build-up of the feedback activity that was observed exclusively for high complexity scenes. Together, these results suggest that while feed-forward activity may suffice for simple scenes, the brain employs recurrent processing more adaptively in naturalistic settings, using minimal feedback for sparse, coherent scenes and increasing feedback for complex, fragmented scenes. Author summary How much neural processing is required to detect objects of interest in natural scenes? The astonishing speed of object recognition suggests that fast feed-forward buildup of perceptual activity is sufficient. However, this view is contradicted by findings that show that disruption of slower neural feedback leads to decreased detection performance. Our study unites these discrepancies by identifying scene complexity as a critical driver of neural feedback. We show how feedback is enhanced for complex, cluttered scenes compared to simple, well-organized scenes. Moreover, for complex scenes, more feedback is associated with better performances. These findings relate the flexibility of neural processes to perceptual decision-making by demonstrating that the brain dynamically directs neural resources based on the complexity of real-world visual inputs.

Similarity judgments and cortical visual responses reflect different properties of object and scene categories in naturalistic images

Numerous factors have been reported to underlie the representation of complex images in high-level human visual cortex, including categories (e.g. faces, objects, scenes), animacy, and real-world size, but the extent to which this organization is reflected in behavioral judgments of real-world stimuli is unclear. Here, we compared representations derived from explicit similarity judgments and ultra-high field (7T) fMRI of human visual cortex for multiple exemplars of a diverse set of naturalistic images from 48 object and scene categories. Behavioral judgements revealed a coarse division between man-made (including humans) and natural (including animals) images, with clear groupings of conceptually-related categories (e.g. transportation, animals), while these conceptual groupings were largely absent in the fMRI representations. Instead, fMRI responses tended to reflect a separation of both human and non-human faces/bodies from all other categories. This pattern yielded a statistically significant, but surprisingly limited correlation between the two representational spaces. Further, comparison of the behavioral and fMRI representational spaces with those derived from the layers of a deep neural network (DNN) showed a strong correspondence with behavior in the top-most layer and with fMRI in the mid-level layers. These results suggest that there is no simple mapping between responses in high-level visual cortex and behavior – each domain reflects different visual properties of the images and responses in high-level visual cortex may correspond to intermediate stages of processing between basic visual features and the conceptual categories that dominate the behavioral response. Significance Statement It is commonly assumed there is a correspondence between behavioral judgments of complex visual stimuli and the response of high-level visual cortex. We directly compared these representations across a diverse set of naturalistic object and scene categories and found a surprisingly and strikingly different representational structure. Further, both types of representation showed good correspondence with a deep neural network, but each correlated most strongly with different layers. These results show that behavioral judgments reflect more conceptual properties and visual cortical fMRI responses capture more general visual features. Collectively, our findings highlight that great care must be taken in mapping the response of visual cortex onto behavior, which clearly reflect different information.

Distinct contributions of functional and deep neural network features to scene representation in brain and behavior

Real-world scenes are rich, heterogeneous stimuli that contain inherent correlations between many visual and semantic features, making it difficult to determine how different scene properties contribute to neural representations. Here, we assessed the unique contributions of three behaviorally relevant feature spaces by a) selecting stimuli for which inherent correlations were minimized a priori and b) partitioning the neural variance attributed to each individual feature space. We found that while scene categorization behavior is best explained by a functional feature space reflecting potential actions in scenes, cortical responses in scene-selective areas are best explained by mid-and high-level layers of computational deep neural network models (DNNs). While other regions of extrastriate cortex represented some functional features, our findings reveal a striking dissociation of functional versus DNN features in their contribution to scene categorization and brain responses, indicating that scene-selective cortex and DNNs represent only a subset of behaviorally relevant scene information.

Characterizing the temporal dynamics of object recognition by deep neural networks: role of depth

Convolutional neural networks (CNNs) have recently emerged as promising models of human vision based on their ability to predict hemodynamic brain responses to visual stimuli measured with functional magnetic resonance imaging (fMRI). However, the degree to which CNNs can predict temporal dynamics of visual object recognition reflected in neural measures with millisecond precision is less understood. Additionally, while deeper CNNs with higher numbers of layers perform better on automated object recognition, it is unclear if this also results into better correlation to brain responses. Here, we examined 1) to what extent CNN layers predict visual evoked responses in the human brain over time and 2) whether deeper CNNs better model brain responses. Specifically, we tested how well CNN architectures with 7 (CNN-7) and 15 (CNN-15) layers predicted electro-encephalography (EEG) responses to several thousands of natural images. Our results show that both CNN architectures correspond to EEG responses in a hierarchical spatio-temporal manner, with lower layers explaining responses early in time at electrodes overlying early visual cortex, and higher layers explaining responses later in time at electrodes overlying lateral-occipital cortex. While the explained variance of neural responses by individual layers did not differ between CNN-7 and CNN-15, combining the representations across layers resulted in improved performance of CNN-15 compared to CNN-7, but only after 150 ms after stimulus-onset. This suggests that CNN representations reflect both early (feed-forward) and late (feedback) stages of visual processing. Overall, our results show that depth of CNNs indeed plays a role in explaining time-resolved EEG responses.