Shepard’s Universal Law of Generalization offered a com-pelling case for the first physics-like law in cognitive sciencethat should hold for all intelligent agents in the universe. Shep-ard’s account is based on a rational Bayesian model of general-ization, providing an answer to the question of why such a lawshould emerge. Extending this account to explain how humansuse multiple examples to make better generalizations requiresan additional assumption, called the size principle: hypothesesthat pick out fewer objects should make a larger contributionto generalization. The degree to which this principle warrantssimilarly law-like status is far from conclusive. Typically, eval-uating this principle has not been straightforward, requiringadditional assumptions. We present a new method for evaluat-ing the size principle that is more direct, and apply this methodto a diverse array of datasets. Our results provide support forthe broad applicability of the size principle.

Large-scale behavioral datasets enable researchers to use com-plex machine learning algorithms to better predict human be-havior, yet this increased predictive power does not always leadto a better understanding of the behavior in question. In thispaper, we outline a data-driven, iterative procedure that allowscognitive scientists to use machine learning to generate mod-els that are both interpretable and accurate. We demonstratethis method in the domain of moral decision-making, wherestandard experimental approaches often identify relevant prin-ciples that influence human judgments, but fail to generalizethese findings to “real world” situations that place these prin-ciples in conflict. The recently released Moral Machine datasetallows us to build a powerful model that can predict the out-comes of these conflicts while remaining simple enough to ex-plain the basis behind human decisions.

Human concept learning is surprisingly robust, allowing forprecise generalizations given only a few positive examples.Bayesian formulations that account for this behavior requireelaborate, pre-specified priors, leaving much of the learningprocess unexplained. More recent models of concept learningbootstrap from deep representations, but the deep neural net-works are themselves trained using millions of positive and neg-ative examples. In machine learning, recent progress in meta-learning has provided large-scale learning algorithms that canlearn new concepts from a few examples, but these approachesstill assume access to implicit negative evidence. In this paper,we formulate a training paradigm that allows a meta-learningalgorithm to solve the problem of concept learning from fewpositive examples. The algorithm discovers a taxonomic prioruseful for learning novel concepts even from held-out supercat-egories and mimics human generalization behavior—the firstto do so without hand-specified domain knowledge or negativeexamples of a novel concept.

Generative models of human identity and appearance havebroad applicability to behavioral science and technology, butthe exquisite sensitivity of human face perception means thattheir utility hinges on the alignment of the model’s representa-tion to human psychological representations and the photoreal-ism of the generated images. Meeting these requirements is anexacting task, and existing models of human identity and ap-pearance are often unworkably abstract, artificial, uncanny, orbiased. Here, we use a variational autoencoder with an autore-gressive decoder to learn a face space from a uniquely diversedataset of portraits that control much of the variation irrele-vant to human identity and appearance. Our method generatesphotorealistic portraits of fictive identities with a smooth, navi-gable latent space. We validate our model’s alignment with hu-man sensitivities by introducing a psychophysical Turing testfor images, which humans mostly fail. Lastly, we demonstratean initial application of our model to the problem of fast searchin mental space to obtain detailed “police sketches” in a smallnumber of trials.

Vector-space representations provide geometric tools for rea-soning about the similarity of a set of objects and their relation-ships. Recent machine learning methods for deriving vector-space embeddings of words (e.g., word2vec) have achievedconsiderable success in natural language processing. Thesevector spaces have also been shown to exhibit a surprising ca-pacity to capture verbal analogies, with similar results for nat-ural images, giving new life to a classic model of analogies asparallelograms that was first proposed by cognitive scientists.We evaluate the parallelogram model of analogy as applied tomodern word embeddings, providing a detailed analysis of theextent to which this approach captures human relational sim-ilarity judgments in a large benchmark dataset. We find thatthat some semantic relationships are better captured than oth-ers. We then provide evidence for deeper limitations of the par-allelogram model based on the intrinsic geometric constraintsof vector spaces, paralleling classic results for first-order simi-larity.

Understanding how people represent categories is a core prob-lem in cognitive science. Decades of research have yieldeda variety of formal theories of categories, but validating themwith naturalistic stimuli is difficult. The challenge is that hu-man category representations cannot be directly observed andrunning informative experiments with naturalistic stimuli suchas images requires a workable representation of these stimuli.Deep neural networks have recently been successful in solvinga range of computer vision tasks and provide a way to com-pactly represent image features. Here, we introduce a methodto estimate the structure of human categories that combinesideas from cognitive science and machine learning, blendinghuman-based algorithms with state-of-the-art deep image gen-erators. We provide qualitative and quantitative results as aproof-of-concept for the method’s feasibility. Samples drawnfrom human distributions rival those from state-of-the-art gen-erative models in quality and outperform alternative methodsfor estimating the structure of human categories.

Traditional models of category learning in psychology focuson representation at the category level as opposed to the stim-ulus level, even though the two are likely to interact. Thestimulus representations employed in such models are eitherhand-designed by the experimenter, inferred circuitously fromhuman judgments, or borrowed from pretrained deep neuralnetworks that are themselves competing models of categorylearning. In this work, we extend classic prototype and ex-emplar models to learn both stimulus and category represen-tations jointly from raw input. This new class of models canbe parameterized by deep neural networks (DNN) and trainedend-to-end. Following their namesakes, we refer to them asDeep Prototype Models, Deep Exemplar Models, and DeepGaussian Mixture Models. Compared to typical DNNs, wefind that their cognitively inspired counterparts both providebetter intrinsic fit to human behavior and improve ground-truthclassification.