Computational models of semantics have emerged as powerful tools for natural language processing. Recent work has developed models to handle compositionality, but these models have typically been evaluated on large, uncontrolled corpora. In this paper, we constructed a controlled set of phrase pairs and collected phrase similarity judgments, revealing novel insights into hu- man semantic representation. None of the computa- tional models that we considered were able to capture the pattern of human judgments. The results of a sec- ond experiment, using the same stimuli with a trans- formational judgment task, support a transformational account of similarity, according to which the similarity between phrases is inversely related to the number of ed- its required to transform one mental model into another. Taken together, our results indicate that popular mod- els of compositional semantics do not capture important facets of human semantic representation.

From social networks to public transportation, graph structuresare a ubiquitous feature of life. How do humans learn functionson graphs, where relationships are defined by the connectiv-ity structure? We adapt a Bayesian framework for functionlearning to graph structures, and propose that people performgeneralization by assuming that the observed function valuesdiffuse across the graph. We evaluate this model by askingparticipants to make predictions about passenger volume in avirtual subway network. The model captures both generaliza-tion and confidence judgments, and provides a quantitativelysuperior account relative to several heuristic models. Our worksuggests that people exploit graph structure to make general-izations about functions in complex discrete spaces.

Bayesian models of cognition posit that people compute prob-ability distributions over hypotheses, possibly by construct-ing a sample-based approximation. Since people encountermany closely related distributions, a computationally efficientstrategy is to selectively reuse computations – either the sam-ples themselves or some summary statistic. We refer to thesereuse strategies as amortized inference. In two experiments,we present evidence consistent with amortization. When se-quentially answering two related queries about natural scenes,we show that answers to the second query vary systematicallydepending on the structure of the first query. Using a cog-nitive load manipulation, we find evidence that people cachesummary statistics rather than raw sample sets. These resultsenrich our notions of how the brain approximates probabilisticinference.

An important challenge for human-like AI is compositional se-mantics. Recent research has attempted to address this by us-ing deep neural networks to learn vector space embeddings ofsentences, which then serve as input to other tasks. We presenta new dataset for one such task, “natural language inference”(NLI), that cannot be solved using only word-level knowledgeand requires some compositionality. We find that the perfor-mance of state of the art sentence embeddings (InferSent; Con-neau et al., 2017) on our new dataset is poor. We analyzethe decision rules learned by InferSent and find that they arelargely driven by simple heuristics that are ecologically validin its training dataset. Further, we find that augmenting train-ing with our dataset improves test performance on our datasetwithout loss of performance on the original training dataset.This highlights the importance of structured datasets in betterunderstanding and improving AI systems.

In an optimal stopping problem, people encounter a sequenceof options and are tasked with choosing the best one; once anoption is rejected, it is no longer available. Recent studies ofoptimal stopping suggest that people compare the current op-tion with an internal threshold and accept it when the optionexceeds the threshold. In contrast, we propose that humans de-cide to accept or reject an option based on an estimate of theprobability that a better option will be observed in the future.We develop a computational model that formalizes this idea,and compare the model to the optimal policy in two experi-ments. Our model provides a better account of the data thanthe optimal model. In particular, our model explains how thedistributional structure of option values affects stopping behav-ior, providing a step towards a more complete psychologicaltheory of optimal stopping.

Models of concept learning and theory acquisition often in-voke a stochastic search process, in which learners generatehypotheses through some structured random process and thenevaluate them on some data measuring their quality or value.To be successful within a reasonable time-frame, these mod-els need ways of generating good candidate hypotheses evenbefore the data are considered. Schulz (2012a) has proposedthat studying the origins of new ideas in more everyday con-texts, such as how we think up new names for things, can pro-vide insight into the cognitive processes that generate good hy-potheses for learning. We propose a simple generative modelfor how people might draw on their experience to proposenew names in everyday domains such as pub names or actionmovies, and show that it captures surprisingly well the namesthat people actually imagine. We discuss the role for an anal-ogous hypothesis-generation mechanism in enabling and con-straining causal theory learning.

How do we perceive the predictability of functions? We derive a rational measure of a function's predictabil- ity based on Gaussian process learning curves. Using this measure, we show that the smoothness of a func- tion can be more important to predictability judgments than the variance of additive noise or the number of samples. These patterns can be captured well by the learning curve for Gaussian process regression, which in turn crucially depends on the eigenvalue spectrum of the covariance function. Using approximate bounds on the learning curve, we model participants' predictabil- ity judgments about sampled functions and ?nd that smoothness is indeed a better predictor for perceived predictability than both the variance and the sample size. This means that it can sometimes be preferable to learn about noisy but smooth functions instead of deterministic complex ones.

Dopamine release in the nucleus accumbens has been hypothesized to signal reward prediction error, the difference between observed and predicted reward, suggesting a biological implementation for reinforcement learning. Rigorous tests of this hypothesis require assumptions about how the brain maps sensory signals to reward predictions, yet this mapping is still poorly understood. In particular, the mapping is non-trivial when sensory signals provide ambiguous information about the hidden state of the environment. Previous work using classical conditioning tasks has suggested that reward predictions are generated conditional on probabilistic beliefs about the hidden state, such that dopamine implicitly reflects these beliefs. Here we test this hypothesis in the context of an instrumental task (a two-armed bandit), where the hidden state switches repeatedly. We measured choice behavior and recorded dLight signals reflecting dopamine release in the nucleus accumbens core. Model comparison based on the behavioral data favored models that used Bayesian updating of probabilistic beliefs. These same models also quantitatively matched the dopamine measurements better than non-Bayesian alternatives. We conclude that probabilistic belief computation plays a fundamental role in instrumental performance and associated mesolimbic dopamine signaling.