Preexisting Knowledge Versus On-Line Learning

Spatial relation categorization in infants and deep neural networks

Spatial relations, such as above, below, between, and containment, are important mediators in children's understanding of the world (Piaget, 1954). The development of these relational categories in infancy has been extensively studied (Quinn, 2003) yet little is known about their computational underpinnings. Using developmental tests, we examine the extent to which deep neural networks, pretrained on a standard vision benchmark or egocentric video captured from one baby's perspective, form categorical representations for visual stimuli depicting relations. Notably, the networks did not receive any explicit training on relations. We then analyze whether these networks recover similar patterns to ones identified in development, such as reproducing the relative difficulty of categorizing different spatial relations and different stimulus abstractions. We find that the networks we evaluate tend to recover many of the patterns observed with the simpler relations of "above versus below" or "between versus outside", but struggle to match developmental findings related to "containment". We identify factors in the choice of model architecture, pretraining data, and experimental design that contribute to the extent the networks match developmental patterns, and highlight experimental predictions made by our modeling results. Our results open the door to modeling infants' earliest categorization abilities with modern machine learning tools and demonstrate the utility and productivity of this approach.

A language of thought for the mental representation of geometric shapes

In various cultures and at all spatial scales, humans produce a rich complexity of geometric shapes such as lines, circles or spirals. Here, we propose that humans possess a language of thought for geometric shapes that can produce line drawings as recursive combinations of a minimal set of geometric primitives. We present a programming language, similar to Logo, that combines discrete numbers and continuous integration to form higher-level structures based on repetition, concatenation and embedding, and we show that the simplest programs in this language generate the fundamental geometric shapes observed in human cultures. On the perceptual side, we propose that shape perception in humans involves searching for the shortest program that correctly draws the image (program induction). A consequence of this framework is that the mental difficulty of remembering a shape should depend on its minimum description length (MDL) in the proposed language. In two experiments, we show that encoding and processing of geometric shapes is well predicted by MDL. Furthermore, our hypotheses predict additive laws for the psychological complexity of repeated, concatenated or embedded shapes, which we confirm experimentally.

How do the object-file and physical-reasoning systems interact? Evidence from priming effects with object arrays or novel labels

How do infants reason about simple physical events such as containment, tube, and support events? According to the two-system model, two cognitive systems, the object-file (OF) and physical-reasoning (PR) systems, work together to guide infants' responses to these events. When an event begins, the OF system sends categorical information about the objects and their arrangements to the PR system. This system then categorizes the event, assigns event roles to the objects, and taps the OF system for information about features previously identified as causally relevant for the event category selected. All of the categorical and featural information included in the event's representation is interpreted by the PR system's domain knowledge, which includes core principles such as persistence and gravity. The present research tested a novel prediction of the model: If the OF system could be primed to also send, at the beginning of an event, information about an as-yet-unidentified feature, the PR system would then interpret this information using its core principles, allowing infants to detect core violations involving the feature earlier than they normally would. We examined this prediction using two types of priming manipulations directed at the OF system, object arrays and novel labels. In six experiments, infants aged 7-13 months (N = 304) were tested using different event categories and as-yet-unidentified features (color in containment events, height in tube events, and proportional distribution in support events) as well as different tasks (violation-of-expectation and action tasks). In each case, infants who were effectively primed reasoned successfully about the as-yet-unidentified feature, sometimes as early as six months before they would typically do so. These converging results provide strong support for the two-system model and for the claim that uncovering how the OF and PR systems represent and exchange information is essential for understanding how infants respond to physical events.

Hand size representation in healthy children and young adults

Whereas we experience our body as a coherent volumetric object, the brain appears to maintain highly fragmented representations of individual body parts. Little is known about how body representations of hand size and shape are built and evolve during infancy and young adulthood. This study aimed to investigate the effect of hand side, handedness, and age on the development of central hand size representation. The observational study with comparison groups was conducted with 90 typically developing Belgian school children and young adults (48 male and 42 female; age range = 5.0-23.0 years; 49 left-handed and 41 right-handed). Participants estimated their hand size and shape using two different tasks. In the localization task, participants were verbally cued to judge the locations of 10 anatomical landmarks of an occluded hand. An implicit hand size map was constructed and compared with actual hand dimensions. In the template selection task, the explicit hand shape was measured with a depictive method. Hand shape indexes were calculated and compared for the actual, implicit, and explicit conditions. Participants were divided into four age groups (5-8 years, 9-10 years, 11-16 years, and 17-23 years). Implicit hand maps featured underestimation of finger length and overestimation of hand width, which is already present in the youngest children. Linear mixed modeling revealed no influence of hand side on finger length underestimation; nonetheless, a significant main effect of age (p = .001) was exposed. Sinistrals aged 11 to 16 years showed significantly less underestimation (p = .03) than dextrals of the same age. As for the hand shape, the implicit condition differed significantly with the actual and explicit conditions (p < .001). Again, the implicit shape index was subjected to handedness and age effects, with significant differences being found between sinistrals and dextrals in the age groups of 9 and 10 years (p = .029) and 11 to 16 years (p < .001). In conclusion, the implicit metric component of the hand representation in children and young adults is misperceived, featuring shortened fingers and broadened hands since a very young age. Crucially, the finger length underestimation increases with age and shows a different developmental trajectory for sinistrals and dextrals. In contrast, the explicit hand shape is approximately veridical and seems immune from age and handedness effects. This study confirms the dual character of somatoperception and establishes a point of reference for children and young adults.

Regularity detection and explanation-based learning jointly support learning about physical events in early infancy

The present research considers statistical learning (SL) and explanation-based learning (EBL) as joint mechanisms to support the development of physical knowledge. Infants watched teaching events in which a cover was lowered over an object and released, with outcomes that violated object principles. The object became fully hidden under a cover that was much shorter, and it remained partly visible under a cover that was much taller. Next, infants watched two test events identical to the teaching events except that one of the events was modified to present a plausible outcome and thus deviated from teaching. Infants at 3.5 months readily detected the regularity in the teaching events and noticed the change in the modified test event, whereas 6.5-month-olds did not. The pattern of response was reversed (1) when 3.5-month-olds were primed to notice the violation of object principles in the teaching events, which interfered with EBL and led infants to miss the change in the modified test event; and (2) when 6.5-month-olds were provided ways to remove the violation from the teaching events, which enabled EBL and led infants to notice the change in the modified test event. Together, the results shed light on young infants' approach to learning about physical events-one that integrates SL for pattern detection and EBL for causal coherence of the rule being learned.

Thinking about quantity: the intertwined development of spatial and numerical cognition

There are many continuous quantitative dimensions in the physical world. Philosophical, psychological, and neural work has focused mostly on space and number. However, there are other important continuous dimensions (e.g., time and mass). Moreover, space can be broken down into more specific dimensions (e.g., length, area, and density) and number can be conceptualized discretely or continuously (i.e., natural vs real numbers). Variation on these quantitative dimensions is typically correlated, e.g., larger objects often weigh more than smaller ones. Number is a distinctive continuous dimension because the natural numbers (i.e., positive integers) are used to quantify collections of discrete objects. This aspect of number is emphasized by teaching of the count word sequence and arithmetic during the early school years. We review research on spatial and numerical estimation, and argue that a generalized magnitude system is the starting point for development in both domains. Development occurs along several lines: (1) changes in capacity, durability, and precision, (2) differentiation of the generalized magnitude system into separable dimensions, (3) formation of a discrete number system, i.e., the positive integers, (4) mapping the positive integers onto the continuous number line, and (5) acquiring abstract knowledge of the relations between pairs of systems. We discuss implications of this approach for teaching various topics in mathematics, including scaling, measurement, proportional reasoning, and fractions.

25 years of research on the use of geometry in spatial reorientation: a current theoretical perspective

The purpose of this article is to review and evaluate the range of theories proposed to explain findings on the use of geometry in reorientation. We consider five key approaches and models associated with them and, in the course of reviewing each approach, five key issues. First, we take up modularity theory itself, as recently revised by Lee and Spelke (Cognitive Psychology, 61, 152-176, 2010a; Experimental Brain Research, 206, 179-188, 2010b). In this context, we discuss issues concerning the basic distinction between geometry and features. Second, we review the view-matching approach (Stürzl, Cheung, Cheng, & Zeil, Journal of Experimental Psychology: Animal Behavior Processes, 34, 1-14, 2008). In this context, we highlight the possibility of cross-species differences, as well as commonalities. Third, we review an associative theory (Miller & Shettleworth, Journal of Experimental Psychology: Animal Behavior Processes, 33, 191-212, 2007; Journal of Experimental Psychology: Animal Behavior Processes, 34, 419-422, 2008). In this context, we focus on phenomena of cue competition. Fourth, we take up adaptive combination theory (Newcombe & Huttenlocher, 2006). In this context, we focus on discussing development and the effects of experience. Fifth, we examine various neurally based approaches, including frameworks proposed by Doeller and Burgess (Proceedings of the National Academy of Sciences of the United States of America, 105, 5909-5914, 2008; Doeller, King, & Burgess, Proceedings of the National Academy of Sciences of the United States of America, 105, 5915-5920, 2008) and by Sheynikhovich, Chavarriaga, Strösslin, Arleo, and Gerstner (Psychological Review, 116, 540-566, 2009). In this context, we examine the issue of the neural substrates of spatial navigation. We conclude that none of these approaches can account for all of the known phenomena concerning the use of geometry in reorientation and clarify what the challenges are for each approach.

Twenty-Five Years Using the Intermodal Preferential Looking Paradigm to Study Language Acquisition

The intermodal preferential looking paradigm (IPLP) has proven to be a revolutionary method for the examination of infants’ emerging language knowledge. In the IPLP, infants’ language comprehension is measured by their differential visual fixation to two images presented side-by-side when only one of the images matches an accompanying linguistic stimulus. Researchers can examine burgeoning knowledge in the areas of phonology, semantics, syntax, and morphology in infants not yet speaking. The IPLP enables the exploration of the underlying mechanisms involved in language learning and illuminates how infants identify the correspondences between language and referents in the world. It has also fostered the study of infants’ conceptions of the dynamic events that language will express. Exemplifying translational science, the IPLP is now being investigated for its clinical and diagnostic value.

Doing Socrates experiment right: controlled rearing studies of geometrical knowledge in animals

The issue of whether encoding of geometric information for navigational purposes crucially depends on environmental experience or whether it is innately predisposed in the brain has been recently addressed in controlled rearing studies. Non-human animals can make use of the geometric shape of an environment for spatial reorientation and in some circumstances reliance on purely geometric information (metric properties and sense) can overcome use of local featural information. Animals reared in home cages of different geometric shapes proved to be equally capable of learning and performing navigational tasks based on geometric information. The findings suggest that effective use of geometric information for spatial reorientation does not require experience in environments with right angles and metrically distinct surfaces.

Can infants be "taught" to attend to a new physical variable in an event category? The case of height in covering events

As they observe or produce events, infants identify variables that help them predict outcomes in each category of events. How do infants identify a new variable? An explanation-based learning (EBL) account suggests three essential steps: (1) observing contrastive outcomes relevant to the variable; (2) discovering the conditions associated with these outcomes; and (3) generating an explanation for the condition-outcome regularity discovered. In Experiments 1-3, 9-month-old infants watched events designed to "teach" them the variable height in covering events. After watching these events, designed in accord with the EBL account, the infants detected a height violation in a covering event, three months earlier than they ordinarily would have. In Experiments 4-6, the "teaching" events were modified to remove one of the EBL steps, and the infants no longer detected the height violation. The present findings thus support the EBL account and help specify the processes by which infants acquire their physical knowledge.

Is there an innate geometric module? Effects of experience with angular geometric cues on spatial re-orientation based on the shape of the environment

Non-human animals and human children can make use of the geometric shape of an environment for spatial reorientation and in some circumstances reliance on purely geometric information (metric properties of surfaces and sense) can overcome the use of local featural cues. Little is known as to whether the use of geometric information is in some way reliant on past experience or, as would likely be argued by advocates of the notion of a geometric module, it is innate. We tested the navigational abilities of newborn domestic chicks reared in either rectangular or circular cages. Chicks were trained in a rectangular-shaped enclosure with panels placed at the corners to provide salient featural cues. Rectangular-reared and circular-reared chicks proved equally able to learn the task. When tested after removal of the featural cues, both rectangular- and circular-reared chicks showed evidence that they had spontaneously encoded geometric information. Moreover, when trained in a rectangular-shaped enclosure without any featural cues, chicks reared in rectangular-, circular-, or c-shaped cages proved to be equally able to learn and perform the task using geometric information. These results suggest that effective use of geometric information for spatial reorientation does not require experience in environments with right angles and metrically distinct surfaces, thus supporting the hypothesis of a predisposed geometric module in the animal brain.

Wayfinding Behavior and Spatial Knowledge of Adults and Children in a Virtual Environment

This study investigated the effect of different organizations of landmark-location pairings as fine-space information on wayfinding behavior and spatial knowledge on a total of 90 participants: 30 second graders, 30 sixth graders, and 30 adults. All participants had to find their way to a goal in a virtual environment with either randomized or categorical landmarks, or without any landmarks. Thereafter, they had to find the shortest way from the start position to the goal in two consecutive trials (wayfinding performance), and they had to solve a number of spatial knowledge tasks. The results showed that independent of their categorical function, the existence of landmarks influenced the wayfinding performance of adults and children in the same way. Whereas the presence of landmarks had no effect on spatial survey knowledge, landmark knowledge itself was influenced by the categorical function of the landmarks presented. Moreover, second graders showed limited achievement compared to adults independent of the existence of landmarks. The main results implicate firstly that children at school age indeed are able to use landmark-location pairings as fine-space information like adults during learning an unknown environmental space, and secondly that a dissociation between wayfinding behavior and spatial knowledge might exist.

Inducing infants to detect a physical violation in a single trial

There is increasing evidence that infants' representations of physical events can be enhanced through appropriate experiences in the laboratory. Most of this research has involved administering infants multiple training trials, often with multiple objects. In the present research, 8-month-olds were induced to detect a physical violation in a single trial. The experiments built on previous evidence that for occlusion events, infants encode height information at about age 3.5 months, but for covering events, they encode height information only at about age 12 months. In two experiments, a short cover was first placed in front of a short or a tall object (occlusion event); next, the cover was lowered over the tall object until it became fully hidden (covering event). Exposure to the occlusion event (but not other events in which height information was not encoded) enabled the infants to detect the violation in the covering event, much earlier than they would have otherwise.