Learning same and different relations: cross-species comparisons

How can I find what I want? Can children, chimpanzees and capuchin monkeys form abstract representations to guide their behavior in a sampling task?

concepts are a powerful tool for making wide-ranging predictions in new situations based on little experience. Whereas looking-time studies suggest an early emergence of this ability in human infancy, other paradigms like the relational match to sample task often fail to detect abstract concepts until late preschool years. Similarly, non-human animals show difficulties and often succeed only after long training regimes. Given the considerable influence of slight task modifications, the conclusiveness of these findings for the development and phylogenetic distribution of abstract reasoning is debated. Here, we tested the abilities of 3 to 5-year-old children, chimpanzees, and capuchin monkeys in a unified and more ecologically valid task design based on the concept of "overhypotheses" (Goodman, 1955). Participants sampled high- and low-valued items from containers that either each offered items of uniform value or a mix of high- and low-valued items. In a test situation, participants should switch away earlier from a container offering low-valued items when they learned that, in general, items within a container are of the same type, but should stay longer if they formed the overhypothesis that containers bear a mix of types. We compared each species' performance to the predictions of a probabilistic hierarchical Bayesian model forming overhypotheses at a first and second level of abstraction, adapted to each species' reward preferences. Children and, to a more limited extent, chimpanzees demonstrated their sensitivity to abstract patterns in the evidence. In contrast, capuchin monkeys did not exhibit conclusive evidence for the ability of abstract knowledge formation.

The Perception of Similarity, Difference and Opposition

After considering the pervasiveness of same/different relationships in Psychology and the experimental evidence of their perceptual foundation in Psychophysics and Infant and Comparative Psychology, this paper develops its main argument. Similarity and diversity do not complete the panorama since opposition constitutes a third relationship which is distinct from the other two. There is evidence of this in the previous literature investigating the perceptual basis of opposition and in the results of the two new studies presented in this paper. In these studies, the participants were asked to indicate to what extent pairs of simple bi-dimensional figures appeared to be similar, different or opposite to each other. A rating task was used in Study 1 and a pair comparison task was used in Study 2. Three main results consistently emerged: Firstly, opposition is distinct from similarity and difference which, conversely, are in a strictly inverse relationship. Secondly, opposition is specifically linked to something which points in an allocentrically opposite direction. Thirdly, alterations to the shape of an object are usually associated with the perception of diversity rather than opposition. The implications of a shift from a dyadic (same/different) to a triadic (similar/different/opposite) paradigm are discussed in the final section.

Food-caching mountain chickadees can learn abstract rules to solve a complex spatial-temporal pattern

The use of abstract rules in behavioral decisions is considered evidence of executive functions associated with higher-level cognition. Laboratory studies across taxa have shown that animals may be capable of learning abstract concepts, such as the relationships between items, but often use simpler cognitive abilities to solve tasks. Little is known about whether or how animals learn and use abstract rules in natural environments. Here, we tested whether wild, food-caching mountain chickadees (Poecile gambeli) could learn an abstract rule in a spatial-temporal task in which the location of a food reward rotated daily around an 8-feeder square spatial array for up to 34 days. Chickadees initially searched for the daily food reward by visiting the most recently rewarding locations and then moving backward to visit previously rewarding feeders, using memory of previous locations. But by the end of the task, chickadees were more likely to search forward in the correct direction of rotation, moving away from the previously rewarding feeders. These results suggest that chickadees learned the direction rule for daily feeder rotation and used this to guide their decisions while searching for a food reward. Thus, chickadees appear to use an executive function to make decisions on a foraging-based task in the wild. VIDEO ABSTRACT.

Over-reliance on English hinders cognitive science

English is the dominant language in the study of human cognition and behavior: the individuals studied by cognitive scientists, as well as most of the scientists themselves, are frequently English speakers. However, English differs from other languages in ways that have consequences for the whole of the cognitive sciences, reaching far beyond the study of language itself. Here, we review an emerging body of evidence that highlights how the particular characteristics of English and the linguistic habits of English speakers bias the field by both warping research programs (e.g., overemphasizing features and mechanisms present in English over others) and overgeneralizing observations from English speakers' behaviors, brains, and cognition to our entire species. We propose mitigating strategies that could help avoid some of these pitfalls.

Can deep convolutional neural networks support relational reasoning in the same-different task?

Same-different visual reasoning is a basic skill central to abstract combinatorial thought. This fact has lead neural networks researchers to test same-different classification on deep convolutional neural networks (DCNNs), which has resulted in a controversy regarding whether this skill is within the capacity of these models. However, most tests of same-different classification rely on testing on images that come from the same pixel-level distribution as the training images, yielding the results inconclusive. In this study, we tested relational same-different reasoning in DCNNs. In a series of simulations we show that models based on the ResNet architecture are capable of visual same-different classification, but only when the test images are similar to the training images at the pixel level. In contrast, when there is a shift in the testing distribution that does not change the relation between the objects in the image, the performance of DCNNs decreases substantially. This finding is true even when the DCNNs’ training regime is expanded to include images taken from a wide range of different pixel-level distributions or when the model is trained on the testing distribution but on a different task in a multitask learning context. Furthermore, we show that the relation network, a deep learning architecture specifically designed to tackle visual relational reasoning problems, suffers the same kind of limitations. Overall, the results of this study suggest that learning same-different relations is beyond the scope of current DCNNs.

Relational thinking: An overlooked component of executive functioning

Relational thinking, the ability to represent abstract, generalizable relations, is a core component of reasoning and human cognition. Relational thinking contributes to fluid reasoning and academic achievement, particularly in the domain of math. However, due to the complex nature of many fluid reasoning tasks, it has been difficult to determine the degree to which relational thinking has a separable role from the cognitive processes collectively known as executive functions (EFs). Here, we used a simplified reasoning task to better understand how relational thinking contributes to math achievement in a large, diverse sample of elementary and middle school students (N = 942). Students also performed a set of ten adaptive EF assessments, as well as tests of math fluency and fraction magnitude comparison. We found that relational thinking was significantly correlated with each of the three EF composite scores previously derived from this dataset, albeit no more strongly than they were with each other. Further, relational thinking predicted unique variance in students' math fluency and fraction magnitude comparison scores over and above the three EF composites. Thus, we propose that relational thinking be considered an EF in its own right as one of the core, mid-level cognitive abilities that supports cognition and goal-directed behavior. RESEARCH HIGHLIGHTS: Relational thinking, the process of identifying and integrating relations, develops over childhood and is central to reasoning. We collected data from nearly 1000 elementary and middle schoolers on a test of relational thinking, ten standard executive function tasks, and two math tests. Relational thinking predicts unique variance in math achievement not accounted for by canonical EFs throughout middle childhood. We propose that relational thinking should be conceptualized as a core executive function that supports cognitive development and learning.

Understanding the Computational Demands Underlying Visual Reasoning

Visual understanding requires comprehending complex visual relations between objects within a scene. Here, we seek to characterize the computational demands for abstract visual reasoning. We do this by systematically assessing the ability of modern deep convolutional neural networks (CNNs) to learn to solve the synthetic visual reasoning test (SVRT) challenge, a collection of 23 visual reasoning problems. Our analysis reveals a novel taxonomy of visual reasoning tasks, which can be primarily explained by both the type of relations (same-different versus spatial-relation judgments) and the number of relations used to compose the underlying rules. Prior cognitive neuroscience work suggests that attention plays a key role in humans' visual reasoning ability. To test this hypothesis, we extended the CNNs with spatial and feature-based attention mechanisms. In a second series of experiments, we evaluated the ability of these attention networks to learn to solve the SVRT challenge and found the resulting architectures to be much more efficient at solving the hardest of these visual reasoning tasks. Most important, the corresponding improvements on individual tasks partially explained our novel taxonomy. Overall, this work provides a granular computational account of visual reasoning and yields testable neuroscience predictions regarding the differential need for feature-based versus spatial attention depending on the type of visual reasoning problem.

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.

Representations of Abstract Relations in Infancy

relations are considered the pinnacle of human cognition, allowing for analogical and logical reasoning, and possibly setting humans apart from other animal species. Recent experimental evidence showed that infants are capable of representing the abstract relations same and different, prompting the question of the format of such representations. In a propositional language of thought, abstract relations would be represented in the form of discrete symbols. Is this format available to pre-lexical infants? We report six experiments (N = 192) relying on pupillometry and investigating how preverbal 10- to 12-month-old infants represent the relation same. We found that infants' ability to represent the relation same is impacted by the number of individual entities taking part in the relation. Infants could represent that four syllables were the same and generalized that relation to novel sequences (Experiments 1 and 4). However, they failed to generalize the relation same when it involved 5 or 6 syllables (Experiments 2-3), showing that infants' representation of the relation same is constrained by the limits of working memory capacity. Infants also failed to form a representation equivalent to all the same, which could apply to a varying number of same syllables (Experiments 5-6). These results highlight important discontinuities along cognitive development. Contrary to adults, preverbal infants lack a discrete symbol for the relation same, and rather build a representation of the relation by assembling symbols for individual entities.

No evidence for language benefits in infant relational learning

Recent studies have found that infants show relational learning in the first year. Like older children, they can abstract relations such as same or different across a series of exemplars. For older children, language has a major impact on relational learning: labeling a shared relation facilitates learning, while labeling component objects can disrupt learning. Here we ask: Does language influence relational learning at 12 months? Experiment 1 (n = 64) examined the influence of a relational label on learning. Prior to the study, the infants saw three pairs of objects, all labeled "These are same" or "These are different". Experiment 2 (n = 48) examined the influence of object labels prior to the study, with three objects labeled (e.g., "This is a cup, this is a tower."). We compared the present results with those of Ferry et al. (2015), where infants abstracted same and different relations after undergoing a similar paradigm without prior labels. If the effects of language mirror those in older children, we would expect that infants given relational labels (Experiment 1) will be helped in abstracting same and different compared to infants not given labels and that infants given object labels (Experiment 2) will be hindered relative to those not given labels. We found no evidence for either prediction. In Experiment 1, infants who had heard relational labels did not benefit compared to infants who had received no labels (Ferry et al., 2015). In Experiment 2, infants who had heard object labels showed the same patterns as those in Ferry et al. (2015), suggesting that object labels had no effect. This finding is important because it highlights a key difference between the relational learning abilities of infants and those seen in older children, pointing to a protracted process by which language and relational learning become entwined.