Learning

In reality, our understanding of the world is often affected by the interaction with other (natural or artificial) cognitive agents. Personalization algorithms have been discussed as an influential component on the internet that possibly limits the pursuit of accurate knowledge, causing confirmation bias and polarization (e.g., Pariser, 2011). However, there has been no mechanistic explanation of how such personalization affects internal cognitive processes and learned knowledge. In this study, we aim to explain using a model-based approach how the interaction with personalization algorithms can hinder optimal category learning. Here, we assume that an exemplar-based learner experiences a baseline, self-directed, or curated learning sequence, while adapting one’s attention via gradient-based optimization (e.g., Krushcke, 1992; Galdo, Vladimir, & Turner, submitted). In the two experimental conditions (i.e., self-directed, curated), either the learner or the external curation algorithm had control over items and features to be learned. Simulation experiments revealed that a personalized learning process can distort the latent representations of categories and misguide learners’ attention, even when a learner is well-intentioned. This observation was generalizable to different datasets with various underlying structural forms. Lastly, when the learners’ knowledge was tested using an independent test dataset, personalized learners tend to show overconfidence about their decisions compared to their predictive accuracy.

Introduction: We conducted a study (N=100) during the COVID-19 pandemic of learning a complex task (i.e., using multiple types of knowledge), troubleshooting single-fault scenarios on a 51-component hierarchical system based on a real radar over multiple days to measure how learning and retention are affected by the training and retention schedule. The schedule is based on a published ACT-R model presented at ICCM. Method: Participants completed 1, 2, or 4 practice sessions on consecutive days that included a 27-minute session studying the material using an online tutor, a 5-minute practice session troubleshooting the simulation and a declarative and recognition memory assessment about the radar. After the practice session(s), participants returned after a delay for a test session that included a memory assessment, a troubleshooting session, and a working memory test. Delays ranged from 3 to 14 days (9 conditions, N=10/condition). The last of the 403 sessions was 30 April 2021. Results: We expect to see a learning curve and a retention curve on a complex task over relatively long learning and retention periods. We expect to see three different retention curves representing declarative decay, mixed decay, and procedural knowledge decay. The effect of delay between practice and test is expected to be moderated by the number of practice sessions. Conclusion: We will present preliminary data on a complex task learned up to four times with delays up 14 days. Results will be compared with the ACT-R learning and retention equations, including suggestions for changes to the theory.

Network analyses of typical language development indicate that words associated with many other words are acquired earlier, implying that typically developing (TD) children are sensitive to the semantic structure of their environment. Children with autism spectrum disorders (ASD) often lag behind their TD peers with respect to language acquisition, despite relatively spared statistical learning and fast-mapping skills. Recent work indicates that children with ASD may struggle with processing the semantic relationships that are the basis for word meaning. We acquired parent-report vocabulary checklists (Communicative Development Inventory; CDI) for 203 ASD children aged 11 – 173 months from the National Database of Autism Research and for 1,096 vocabulary matched TD children aged 11 – 30 months from WordBank to establish vocabulary composition. To estimate the semantic structure of these vocabularies, we referenced child-oriented word association data to construct an associative network from each child’s vocabulary. Network structure statistics were modeled as a function of group (TD/ASD) and vocabulary size (linear, quadradic, and cubic trends). Network structure developed along different trajectories for each group as vocabularies grew. This began early in vocabulary acquisition, with vocabularies in the ASD group developing clusters more rapidly than the TD group until acquiring about 150 words. After that point, network statistics converged between groups as vocabularies become more similar (i.e., they begin to saturate the CDI wordlist). This suggests that children with ASD have a distinctive trajectory of vocabulary growth that, relative to TD children, is more oriented towards clusters of semantically related words early in language acquisition.

Multi-armed bandits are a useful paradigm to study how people balance exploration (learning about the value of options) and exploitation (choosing options with known high value). When options are distinguished by features predictive of reward, exploration aids generalization of experience to unknown options. The present study builds on our earlier work on human exploration and generalization in a feature-based bandit task (Stojic et al., 2020). Here, I present results from a new experiment where novel options are introduced regularly in three different environments: options either only provide rewards (gain), only provide punishments (loss), or can both provide rewards or punishments (mixed). Options were represented by randomly generated tree-like shapes, with features determining the angle and width of branches. Value of the options was a nonlinear function of the features. Regardless of the environment, people were quite good at choosing the best option. When first encountering each novel option, whether that option was chosen depended on the relative value of the option, indicative of successful function generalization. Compared to the other environments, exploration of novel options was generally larger in the loss environment. Computational modelling provides further insights into these results. We contrast a model that employs function learning through Gaussian Process regression with a new model that learns the value of options through a hierarchical Bayesian filter. Both models can employ a Bayesian mechanism to allow for asymmetric learning rates for positive vs negative reward prediction errors. Some evidence for such asymmetric learning is found.

Two of the most fundamental difficulties we face when learning is deciding which information is relevant, and when to use it. To overcome these difficulties, humans continuously make choices about which dimensions of information to selectively attend to, and monitor how useful those dimensions are in the context of the current goal. Although previous theories have specified how observers learn to attend to relevant dimensions over time, those theories have largely remained silent about how attention should be allocated on a within-trial basis, which dimensions of information should be sampled, and how the temporal ordering of information sampling influences learning. Here, we use the Adaptive Attention and Representation Model (AARM) to demonstrate that a common set of mechanisms can be used to specify: 1) how the distribution of attention is updated between trials over the course of learning; and 2) how attention dynamically shifts among dimensions within-trial. We validate or proposed set of mechanisms by comparing AARM’s predictions to observed behavior in the context of five case studies, which collectively encompass different theoretical aspects of selective attention. Importantly, we use both eye-tracking and choice response data to provide a stringent test of how attention and decision processes dynamically interact. Specifically, how does attention to selected stimulus dimensions gives rise to decision dynamics, and in turn, how do decision dynamics influence our continuous choices about which dimensions to attend to via gaze fixations?

There are considerable differences of opinion about the underlying mechanisms of major depressive disorder. While some emphasize the importance of reward learning, others focus more on a negative mood, and still others emphasize the role of getting stuck in persistent negative thinking. In this task, I will present data from various tasks assessing these different cognitive mechanisms, and show that while impairments in reward learning are associated with depression scores, objective measures of persistent negative thinking are associated with rumination scores. I will then discuss how ACT-R models can explain the effects of persistent negative thinking on task performance in various tasks.