Dr. Sudeep Bhatia
People learn new categories on a daily basis, and the study of category learning is a major topic of research in cognitive science. However, most prior work has focused on how people learn categories over abstracted, artificial (and usually perceptual) representations. Little is known about how new categories are learnt for natural objects, for which people have extensive prior knowledge. We examine this question in three pre-registered studies involving the learning of new categories for everyday foods. Our models use word vectors derived from large-scale natural language data to proxy mental representations for foods, and apply classical models of categorization over these vectorized representations to predict participant categorization judgments. This approach achieves high predictive accuracy rates, and can be used to identify the real-world settings in which category learning is impaired. In doing so, it shows how existing theories of categorization can be used to predict and improve everyday cognition and behavior.
In previous work, we showed how different learning contexts affected not only choice proportion but also decision time: Participants tend to give faster responses in higher-value contexts compared to low-value contexts. To explain these effects, we combined traditional reinforcement learning models–which model across-trial dynamics–with sequential sampling models–which model within-trial dynamics. However, it remains to be assessed whether the magnitude and sign of rewards are associated with different decision mechanisms (i.e., decision caution or motor facilitation). In this study, we manipulated both the magnitude and sign of rewards in a within-participant design. We found that the two manipulations had overall different effects on the joint choice proportion and response times patterns. We propose a new model that attempts to explain such patterns and therefore provide a concise and comprehensive account of value effects on decision-making in reinforcement learning.
Multidimensional scaling (MDS) has provided insight into the structure of human perception and conceptual knowledge and has been central in the development of models of cognition. However, MDS usually requires participants to produce large numbers of similarity judgments, leading to long and repetitive experiments. Here we propose a method that combines a simple grouping task with a neural network model to uncover participants' psychological spaces. We validate the method on simulated data and find that it can recover the true global structure even when given heterogeneous groupings. We then apply the method to data from the World Color Survey and find that it can learn language-specific color organization. Finally, we apply the method to a novel developmental experiment and find age-dependent differences in conceptual spaces. Our results suggest that the method can be used to recover similarity judgments from populations for which traditional MDS setups would be prohibitively taxing, such as in developmental studies. These similarities, in turn, are crucial for the development of detailed models of category learning.
Dr. Olivera Savic
Ms. Alexandria Barkhimer
Dr. Vladimir Sloutsky
Prof. Simon Dennis
Decades of cognitive development research focused on how and when human learners acquire taxonomic links - links that connect concepts belonging to the same semantic category, such as fruit, bird, or furniture. As learning taxonomic links requires the ability to detect key features shared by members of a semantic category, studies have shown that the formation of taxonomic links in semantic memory has a protracted development. However, recent studies report that taxonomic links may be formed even at the age of six months. The goal of the current study was to provide an explanation of the inconsistency across these studies. We examined the possibility that taxonomic links that are acquired early in life will also co-occur frequently, and, therefore, the seemingly taxonomic links early in life are actually driven by co-occurrence statistics. To test this assumption, we selected studies that claim early taxonomic development and studies that claim protracted taxonomic development. Then we calculated the co-occurrence statistics (i.e., cubed pointwise mutual information) between the word pairs using the TASA corpus. Results showed that the studies supporting an early taxonomic development used word pairs that have high co-occurrence statistics, while the studies supporting a late taxonomic development used word pairs that have low co-occurrence statistics. Our results provide evidence that early in development, links between some taxonomically related concepts may stem from co-occurrence regularities.
Dr. Jerome Busemeyer
Quantum probability theory has successfully provided accurate descriptions of behavior in the areas of judgment and decision making, and here we apply the same principles to two category learning tasks, one task using overlapping, information-integration (II) categories the other using overlapping, rule-based (RB) categories. Since II categories lack verbalizable descriptions, unlike RB categories, we assert that an II categorization decision is constructed out of an indefinite state and characterized by quantum probability theory, whereas an RB categorization decision is read out from a definite state and governed by classical probability theory. In our experiment, participants learn to categorize simple, visual stimuli as members of either category S or category K during an acquisition phase, and then rate the likelihood on a scale of 0 to 5 that a stimulus belongs to one category and subsequently perform the same likelihood rating for the other category during a transfer phase. Following the principle of complementarity in quantum theory, we expect the category likelihood ratings to exhibit order effects in the task that employs II categories, but not in the one that uses RB categories. In the task with II categories, we found that the quantum random walk model notably outperforms an analogous Markov random walk model and there are definitive order effects in the likelihood ratings. But in the task with RB categories, we found that the performance gap between the Markov and quantum models is reduced and the order effects in the likelihood ratings are not significant.
Prof. Ben Newell
Several recent studies have shown a positive effect of incentives on effort and attention in menial tasks such as repeated key-presses or counting numbers of objects on a screen (Dellavigna & Pope 2017; Caplin et al. 2020). If processes like effort and attention are modulated by financial incentives, is the same true of higher order cognitive abilities? We replicate a classic category learning experiment that relies on attending to relevant stimulus features to correctly distinguish two groups of objects (Shepard et al. 1961). Different group assignments of the same stimuli varied the difficulty of the task. We compare the learning and test performance between subjects across a wide range of financial incentive levels, to examine how anticipated reward influences rule generation and inductive reasoning. Our preliminary results join with several recent reports showing little to no modulation of learning performance with incentives (van den Berg, Zou, Ma 2020; Enke et al. 2021). The key emerging distinction concerning the relative effectiveness of incentives is between cognitive tasks requiring novel insight and hypothesis discovery versus those that require rote perseverance.