The contrasting shape representations that support object recognition in humans and CNNs

The success of Convolutional Neural Networks (CNNs) in classifying objects has led to a surge of interest in using these systems to understand human vision. Recent studies have argued that when CNNs are trained in the correct learning environment, they can emulate a key property of human vision -- learning to classify objects based on their shape. While showing a shape-bias is indeed a desirable property for any model of human object recognition, it is unclear whether the resulting shape representations learned by these networks are human-like. We explored this question in the context of a well-known observation from psychology showing that humans encode the shape of objects in terms of relations between object features. To check whether this is also true for the representations of CNNs, we ran a series of simulations where we trained CNNs on datasets of novel shapes and tested them on a set of controlled deformations of these shapes. We found that CNNs do not show any enhanced sensitivity to deformations which alter relations between features, even when explicitly trained on such deformations. This behaviour contrasted with human participants in previous studies as well as in a new experiment. We argue that these results are a consequence of a fundamental difference between how humans and CNNs learn to recognise objects: while CNNs select features that allow them to optimally classify the proximal stimulus, humans select features that they infer to be properties of the distal stimulus. This makes human representations more generalisable to novel contexts and tasks.

Integrated pathogen load and dual transcriptome analysis of systemic host-pathogen interactions in severe malaria

The pathogenesis of severe Plasmodium falciparum malaria is incompletely understood. Since the pathogenic stage of the parasite is restricted to blood, dual RNA-sequencing of host and parasite transcripts in blood can reveal their interactions at a systemic scale. Here we identify human and parasite gene expression associated with severe disease features in Gambian children. Differences in parasite load explained up to 99% of differential expression of human genes but only a third of the differential expression of parasite genes. Co-expression analyses showed a remarkable co-regulation of host and parasite genes controlling translation, and host granulopoiesis genes uniquely co-regulated and differentially expressed in severe malaria. Our results indicate that high parasite load is the proximal stimulus for severe P. falciparum malaria, that there is an unappreciated role for many parasite genes in determining virulence, and hint at a molecular arms-race between host and parasite to synthesise protein products.

Proximal Stimulus for Brewster’s Phenomenon is Equivalent to That of Wheatstone’s Stereogram

We examined Brewster’s perceptual phenomenon that is produced by the same binocularly equivalent proximal stimulus as Wheatstone’s stereogram. We then constructed and observed the perceptual effect created by another distal stimulus that also created an equivalent binocular proximal stimulus. All three stimulus situations produced the appearance of a tilted line in the median plane which is consistent with Hering’s rule that the average of the two local signs and binocular eye position determine visual direction. The characteristics of the proximal stimulus are also relevant to the question raised by Helmholtz regarding the empirical vertical horopter.

Corticosteroids Operate as a Switch between Memory Systems

Stress and corticosteroid hormones are known to affect learning and memory processes. In this study, we examined whether stress and corticosteroids are capable of facilitating the switch between multiple memory systems in mice. For this purpose, we designed a task that allowed measurement of nucleus caudate-based stimulus–response and hippocampus-based spatial learning strategies. Naive mice used spatial strategies to locate an exit hole on a circular hole board at a fixed location flagged by a proximal stimulus. When the mice were either stressed or administered corticosterone before the task, 30–50% of the mice switched to a stimulus–response strategy. This switch between learning strategies was accompanied by a rescue of performance, whereas performance declined in the stressed mice that kept using the spatial strategy. Pretreatment with a mineralocorticoid receptor antagonist prevented the switch toward the stimulus–response strategy but led to deterioration of hippocampus-dependent performance. These findings (i) show that corticosteroids promote the transition from spatial to stimulus–response memory systems, (ii) provide evidence that the mineralocorticoid receptor underlies this corticosteroid-mediated switch, and (iii) suggest that a stress-induced switch from hippocampus-based to nucleus caudate-based memory systems can rescue performance.

‘Perceiving the Present’ as a Framework for Ecological Explanations of the Misperception of Projected Angle and Angular Size

An implicit, underlying assumption of most Helmholtzian/Bayesian approaches to perception is the hypothesis that the scene an observer perceives is the probable source of the proximal stimulus. There is, however, a nontrivial latency (on the order of 100 ms) between the time of a proximal stimulus and the time a visual percept is elicited. It seems plausible that it would be advantageous for an observer to have, at any time t, a percept representative of what is out there at that very time t, not a percept of the recent past. If this is so, it implies a modification to the implicit hypothesis underlying most existing probabilistic approaches to perception: the new hypothesis is that, given the proximal stimulus, the scene an observer perceives is the probable scene present at the time of the percept. That is, the hypothesis is that what an observer perceives is not the probable source of the proximal stimulus, but the probable way the probable source will be when the percept actually occurs. A model of an observer's typical movements in the world is developed, and it is shown that projected angles are perceived in a way consistent with the way the probable source will project to the eye after a small time period of forward movement by the observer. The predicted and actual direction of projected-angle misperception is sometimes toward 90° and sometimes away from 90°, depending on whether the probable source angle is lying in a plane parallel or perpendicular to the probable direction of motion, respectively. The perception of angular size for lines in a figure with cues they are lying in a plane perpendicular to the direction of motion is also shown to fit the predictions of the model.

The Reality of Experience: Gibson's Way

This paper considers some first principles that might provide a basis for an objective science of experience (presence or immersion). Dimensions that are considered include classical Newtonian measures of the distal stimulus, changes in neural mechanisms reflecting the proximal stimulus, information theoretic measures of the statistical properties of events, and functional properties related to intentions and abilities. Gibson's ecological framework is suggested as a promising functional approach for defining the reality of experience in relation to the problem of designing virtual environments. This approach emphasizes the tight coordination between perception and action and fixes the measurement coordinate system relative to the capacity for action.

Grouping Based on Phenomenal Similarity of Achromatic Color

It is widely acknowledged that a precondition for the perception of the world of objects and events is an early process of organization, and it has generally been assumed that such organization is based on the Gestalt laws of grouping. However, the stage at which such grouping occurs, whether early or late, is an empirical question. It is demonstrated in two experiments that grouping by similarity of neutral color is based not on similarity of absolute luminance at the level of the proximal stimulus, but on phenomenal similarity of lightness resulting from the achievement of lightness constancy. An alternative explanation of such grouping based on the equivalence of luminance ratios between elements and background is ruled out by appropriate control conditions.

Size Constancy in Children: A New Interpretation

Existing evidence indicates that there are differences between children and adults in size constancy when observation distances are large. Findings are reported which suggest that this phenomenon is based on a difference in the accessing of proximal stimulus information, which, in the case of size, refers to visual angle subtended. Age differences were found when a traditional size constancy task was used, but these differences disappeared when all the comparison objects subtended the same visual angle. Since this finding demonstrates that young children can make accurate size matches, it is suggested that the underconstancy previously reported is not necessarily the result of childrens' inability to use fully certain cues to distance. Rather, the findings suggest that children access proximal stimulus information more spontaneously than do adults.