Learning & Generalization

How are written (printed) words cognitively represented and processed? This question is commonly approached empirically with masked form priming, in which a brief presentation of a prime stimulus precedes the target of a task such as lexical decision. Models of the facilitatory and inhibitory effects of different types of primes typically assume that (1) over perceptual time, the percept changes primarily in signal-to-noise ratio, so that facilitatory similarity effects must arise from cognitive similarity between complete percepts; and (2) associations between letters (or substrings) and words are homogeneous so that inhibitory effects of other vocabulary items must arise from online interactive processes. In contrast, (1) Adelman (2011) modelled facilitatory effects using stochastic accumulation of separable perceptual features, so that “similarity” arises from ambiguity caused by incomplete percepts; and (2) Adelman and Trifonova (2022) modelled inhibitory effects using associative weights learned through a modified Rescorla and Wagner (1972) rule. The former model cannot accommodate any inhibitory effects, and the latter model gives only activation levels (not predictions in milliseconds). I present a model that Integrates these prior models by modifying the representational assumptions of both. The resulting model has superior performance to other models (including its constituents) on the Adelman et al. (2014) benchmark condition means, and accounts for differences in priming due to target properties (frequency, neighborhood size). I also consider inhibitory effects and individual differences.

This is an in-person presentation on July 19, 2026 (10:40 ~ 11:00 EDT).

Learning is driven by prediction errors, yet what counts as a prediction error depends on how the environment is mentally represented. In experiments, researchers typically assume that learners will have a particular representation in mind, but a naive participant must actively construct and revise their mental models of such an environment. We propose that the learning of such representations is influenced by two representational policies, which constitute metacognitive strategies aimed at improving the correspondence between one’s model and the environment: filtering excludes variables deemed irrelevant from the mental model, whereas investing reflects uncertainty-driven exploration of potentially relevant variables. Across two experiments using a novel type of hierarchical multi-armed bandit task, we manipulated whether an initially irrelevant feature later became informative (i.e., a regularity appeared) or whether a previously informative feature lost its predictive value (i.e., a regularity disappeared) midway through the task. We show that participants who filtered out the initially irrelevant feature failed to detect a simple, deterministic, and fully observable regularity when it emerged. Conversely, investing in that feature enabled adaptive insight when it became predictive, but incurred a cost of greater exploratory effort. Strikingly, representational policies were not determined solely by the availability of observational evidence. Many participants continued to invest in a potentially relevant feature despite the absence of supporting evidence. Increasing exploratory costs reduced this persistence and insight prevalence, demonstrating sensitivity to a cost-adaptability trade-off. Together, our findings suggest that early representational choices function as a bottleneck that constrains which learning outcomes become possible – highlighting how representational policies produce variability and path dependence in human learning.

This is an in-person presentation on July 19, 2026 (11:00 ~ 11:20 EDT).

Exemplar models often define probe–exemplar similarity multiplicatively across dimensions, so that exemplar activation is maximal under perfect matches and decreases rapidly as mismatches accumulate. A related consequence is that attention, when placed inside the similarity exponent, primarily changes exemplar activation in the presence of mismatches rather than matches. This contrasts with a frequently observed temporal dynamic in visual search neurophysiology: early spiking increases broadly for neurons whose receptive fields include the stimulus, regardless of whether it is a target or a distractor, followed by increased firing for targets but decreased firing for nontargets. To reconcile this discrepancy, we propose an alternative mechanism that formalizes partial encoding during similarity computation that drives activation. On each probe–exemplar comparison, only a subset of features is assumed to be available; unobserved features are treated as probabilistically imputed under a prior over features implied by stored exemplars. We implement this in a couple of theoretically interesting ways that marginalize over possible expected features through an imputation process. The resulting rule provides a tractable similarity computation that can be integrated into standard exemplar-based categorization models. Simulations show that the modified similarity equation produces a two-stage activation pattern at the exemplar level: under weak attentional selectivity, both matching and mismatching exemplars exhibit elevated activation; under higher selectivity, activation is preserved for matching exemplars while being progressively suppressed for mismatching exemplars. Moreover, the same simulations also recover the relationship between attentional selectivity and response accuracy: predicted accuracy improves with attentional selectivity with gains tapering at higher selectivity. These patterns are robust across a broad parameter range, and therefore, this mechanism provides a principled bridge between exemplar-based categorization and the temporal dynamics of neural selectivity.

This is an in-person presentation on July 19, 2026 (11:20 ~ 11:40 EDT).