Search

IntroductionA remarkable phenomenon in perception is that the visual system spontaneously organizes sets of discrete elements into abstract shape representations. We studied perceptual performance with dot displays to discover what spatial relationships support shape perception.MethodsIn Experiment 1, we tested conditions that lead dot arrays to be perceived as smooth contours vs. having vertices. We found that the perception of a smooth contour vs. a vertex was influenced by spatial relations between dots beyond the three points that define the angle of the point in question. However, there appeared to be a hard boundary around 90° such that any angle 90° or less was perceived as a vertex regardless of the spatial relations of ancillary dots. We hypothesized that dot arrays whose triplets were perceived as smooth curves would be more readily perceived as a unitary object because they can be encoded more economically. In Experiment 2, we generated dot arrays with and without such “vertex triplets” and compared participants’ phenomenological reports of a unified shape with smooth curves vs. shapes with angular corners. Observers gave higher shape ratings for dot arrays from curvilinear shapes. In Experiment 3, we tested shape encoding using a mental rotation task. Participants judged whether two dot arrays were the same or different at five angular differences. Subjects responded reliably faster for displays without vertex triplets, suggesting economical encoding of smooth displays. We followed this up in Experiment 4 using a visual search task. Shapes with and without vertex triplets were embedded in arrays with 25 distractor dots. Participants were asked to detect which display in a 2IFC paradigm contained a shape against a distractor with random dots. Performance was better when the dots were sampled from a smooth shape than when they were sampled from a shape with vertex triplets.Results and discussionThese results suggest that the visual system processes dot arrangements as coherent shapes automatically using precise smoothness constraints. This ability may be a consequence of processes that extract curvature in defining object shape and is consistent with recent theory and evidence suggesting that 2D contour representations are composed of constant curvature primitives.

Deep convolutional neural networks (DCNNs) have attracted considerable interest as useful devices and as possible windows into understanding perception and cognition in biological systems. In earlier work, we showed that DCNNs differ dramatically from human perceivers in that they have no sensitivity to global object shape. Here, we investigated whether those findings are symptomatic of broader limitations of DCNNs regarding the use of relations. We tested learning and generalization of DCNNs (AlexNet and ResNet-50) for several relations involving objects. One involved classifying two shapes in an otherwise empty field as same or different. Another involved enclosure. Every display contained a closed figure among contour noise fragments and one dot; correct responding depended on whether the dot was inside or outside the figure. The third relation we tested involved a classification that depended on which of two polygons had more sides. One polygon always contained a dot, and correct classification of each display depended on whether the polygon with the dot had a greater number of sides. We used DCNNs that had been trained on the ImageNet database, and we used both restricted and unrestricted transfer learning (connection weights at all layers could change with training). For the same-different experiment, there was little restricted transfer learning (82.2%). Generalization tests showed near chance performance for new shapes. Results for enclosure were at chance for restricted transfer learning and somewhat better for unrestricted (74%). Generalization with two new kinds of shapes showed reduced but above-chance performance (≈66%). Follow-up studies indicated that the networks did not access the enclosure relation in their responses. For the relation of more or fewer sides of polygons, DCNNs showed successful learning with polygons having 3–5 sides under unrestricted transfer learning, but showed chance performance in generalization tests with polygons having 6–10 sides. Experiments with human observers showed learning from relatively few examples of all of the relations tested and complete generalization of relational learning to new stimuli. These results using several different relations suggest that DCNNs have crucial limitations that derive from their lack of computations involving abstraction and relational processing of the sort that are fundamental in human perception.

In blast-related mild traumatic brain injury (br-mTBI) little is known about the connections between initial trauma and expression of individual clinical symptoms. Partly due to limitations of current in vitro and in vivo models of br-mTBI, reliable prediction of individual short- and long-term symptoms based on known blast input has not yet been possible. Here we demonstrate a dose-dependent effect of shock wave exposure on C. elegans using shock waves that share physical characteristics with those hypothesized to induce br-mTBI in humans. Increased exposure to shock waves resulted in decreased mean speed of movement while increasing the proportion of worms rendered paralyzed. Recovery of these two behavioral symptoms was observed during increasing post-traumatic waiting periods. Although effects were observed on a population-wide basis, large interindividual variability was present between organisms exposed to the same highly controlled conditions. Reduction of cavitation by exposing worms to shock waves in polyvinyl alcohol resulted in reduced effect, implicating primary blast effects as damaging components in shock wave induced trauma. Growing worms on NGM agar plates led to the same general results in initial shock wave effect in a standard medium, namely dose-dependence and high interindividual variability, as raising worms in liquid cultures. Taken together, these data indicate that reliable prediction of individual clinical symptoms based on known blast input as well as drawing conclusions on blast input from individual clinical symptoms is not feasible in br-mTBI.