Small numerosity advantage for sequential enumeration on RSVP stimuli: an object individuation-based account

Configural relations in humans and deep convolutional neural networks

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.

Enumerating the forest before the trees: The time courses of estimation-based and individuation-based numerical processing

Ensemble perception refers to the ability to report attributes of a group of objects, rather than focusing on only one or a few individuals. An everyday example of ensemble perception is the ability to estimate the numerosity of a large number of items. The time course of ensemble processing, including that of numerical estimation, remains a matter of debate, with some studies arguing for rapid, "preattentive" processing and other studies suggesting that ensemble perception improves with longer presentation durations. We used a forward-simultaneous masking procedure that effectively controls stimulus durations to directly measure the temporal dynamics of ensemble estimation and compared it with more precise enumeration of individual objects. Our main finding was that object individuation within the subitizing range (one to four items) took about 100-150 ms to reach its typical capacity limits, whereas estimation (six or more items) showed a temporal resolution of 50 ms or less. Estimation accuracy did not improve over time. Instead, there was an increasing tendency, with longer effective durations, to underestimate the number of targets for larger set sizes (11-35 items). Overall, the time course of enumeration for one or a few single items was dramatically different from that of estimating numerosity of six or more items. These results are consistent with the idea that the temporal resolution of ensemble processing may be as rapid as, or even faster than, individuation of individual items, and support a basic distinction between the mechanisms underlying exact enumeration of small sets (one to four items) from estimation.

Why Not Just Features? Reconsidering Infants' Behavior in Individuation Tasks

It counts as empirically proven that infants can individuate objects. Object individuation is assumed to be fundamental in the development of infants' ontology within the object-first account. It crucially relies on an object-file (OF) system, representing both spatiotemporal ("where") and categorical ("what") information about objects as solid, cohesive bodies moving continuously in space and time. However, infants' performance in tasks requiring them to use featural information to detect individuation violations appears to be at odds with the object-first account. In such cases, infants do not appear to be able to develop correct expectations about the numerosity of objects. Recently, proponents of the object-first account proposed that these individuation failures result from integration errors between the OF system and an additional physical reasoning system. We are going to argue that the predictions of a feature-based physical-reasoning (PR) system are sufficient for explaining infants' behavior. The striking predictive power of the PR system calls into question the relevance of the OF system and, thereby, challenges the assumption that infants can individuate objects early on.