Canada

Accounts of human and machine concept learning face a fundamental challenge. Some approaches, notably deep learning, frequently achieve human-level performance on specific tasks but lack consistently human-like—i.e. generalizable, composable, and explainable—solutions. Others, including classic symbolic accounts, produce human-like hypotheses but scale poorly. We present a model of learning which is both human-level and human-like. It represents concepts as program-like expressions formed by applying a series of higher-order inferences that iteratively revise preexisting concepts into novel target concepts. Learning seeks the best combination of revisions under a Bayesian score. This model predicts learning behavior in 392 humans over 100 computationally sophisticated concepts more accurately than alternative models (Enumerate, Metagol, RobustFill, Fleet) while using three orders of magnitude less computation. This work shows how humans plausibly construct sophisticated algorithmic representations, a necessity for compelling human-like artificial intelligence.

CogSci 2022

Learning as Programming: Modeling Efficient Search in Human Concept Learning

technical paper

Join us for CogSci2022! Our 44th Annual Conference will be centered around the theme of diversity, with an accompanying range of keynotes and invited speakers who themselves highlight the diversity of our field. Our three symposia are Cognition across Diverse Populations, Cognition within Diverse Societies, and Global Challenges for Cognitive Science. The conference is being offered in a fully hybrid environment for onsite and online attendees.
 
 [**The Program at a Glance can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2022/07/Cogsci-2022-PAAG-Detailed-July-12final-1.pdf)
 [**The Detailed Program can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2022/08/DeArmond-CogSci2022-program-v113-final.pdf)

This conference requires a registration to attend. Please use the link below to secure your ticket. The website will be available to all registered attendees on July 25th. 

[Register to Attend](https://cognitivesciencesociety.org/registration/)

Registration Is Required

The 44th Annual Meeting of the Cognitive Science Society highlighted research on the theme Cognitive Diversity in addition to the full breadth and diversity of research topics offered by the society’s membership. 





When people think about the features of common objects, like scissors, they often spontaneously recall a central feature: scissors cut things. They tend not to recall other features of scissors, e.g., that they have handles. The present paper posits a novel explanation for the behavior: the features people recall first and most often reflect semantic generalizations of kinds. A recent taxonomy of such generalizations suggests that people represent privileged links between kinds and their features known as principled connections (Prasada et al., 2013). Several tests diagnose principled connections: for instance, principled connections reflect norms, so one way to diagnose the presence of a principled connection is to test the acceptability of sentences of the form all normal Xs have feature Y, as in all normal cars have four wheels. We tested whether participants accept generalizations about the normality of features produced in a semantic feature production task (Experiments 1 and 2) as well as self-referential generalizations (Experiment 3). The experiments provided participants with generalizations about features listed first and most often as well as features that people list less frequently. They found that people readily accepted generalizations that diagnose the presence of principled connections. The results corroborate the view that principled connections help people recall the features of conceptual categories.

Frequently produced semantic features reflect principled connections

Explanation generation is ubiquitous, and is shaped by our individual and collective knowledge about a phenomenon. Some phenomena are not easily explained by known explanations and require the creation of original or unknown explanations (How does scientific knowledge grow?), while others are strongly associated with known explanations (What caused the Battle of Dunkirk?). The processes underlying explanation generation when known/unknown have different characterizations, and broad implications for learning and innovation. However, this distinction is not studied in prior work. We hypothesize that unknown (compared to known) explanations will: require relatively more sense-making––manipulation of acquired information; exhibit self-directed information search not captured by simple information foraging models; and result in looser knowledge networks containing semantically dissimilar foraged-for information. We embed these hypotheses into a model framework, and show that it captures the behavior of participants asked to generate explanations in unknown/known conditions using information they forage for on Wikipedia.

The Generation of Original and Known Explanations: How the presence or absence of relevant information influences sense-making and information foraging

When we read research findings, what facilitates consideration of different possible outcomes? 112 Amazon Mechanical Turk Workers read about four psychological studies, and either predicted the results, received the results without explanation, or received the results with a plausible explanation. We hypothesized that 1. receiving results would increase the difficulty of explaining alternative outcomes, and that this would be amplified by receiving explanations, 2. As difficulty explaining alternatives increases, participants would be less surprised by actual results and predict higher likelihood of replication. We did not find the expected differences in difficulty across conditions. However, across all four research studies, surprise was negatively correlated with likelihood of replication, indicating that more surprising results are considered less likely to replicate. We consider reasons why our manipulation did not affect difficulty of explaining alternative outcomes, and discuss implications for how research can best be presented to promote consideration of alternative outcomes.

Alternative Outcomes, Surprise, and Predicting Replication of Psychological Research

People often evaluate explanations by considering various ‘explanatory virtues’, such as an explanation’s simplicity (i.e. the number of unexplained causes referred to). Simplicity has been thought to guide these evaluations by providing a cue to the inputs of Bayesian inference (priors and likelihoods), thus indirectly helping compute the outputs: the posterior probability of an explanation being true. Yet simplicity may also play other, more direct, roles in explanation evaluations. While study 1 supported the mediating role of priors and likelihoods in people’s simplicity preferences, study 1 and 2 found that participants still preferred simpler explanations after statistically controlling for priors and likelihoods (either elicited or provided). These results suggest that simplicity guides explanation evaluations not just by providing cues to the inputs of Bayesian inference, but also by serving as a direct cue to the outputs of these inferences – perhaps providing a simplifying heuristic for these evaluations.

Simplicity beyond probability: Simplicity’s role in evaluating explanations goes beyond providing cues to priors and likelihoods

How do humans explain and cognize visual information? Why do AI explanations in radiology, despite their remarkable accuracy, fail to gain human trust? In a study of 13 radiology practitioners, we found that AI explanations of x-rays differ from human explanations in 3 ways. The first concerns visual reasoning and evidence: how humans get other humans to see an interpretation’s validity. Machine learned classifications lack this evidentiary grounding, and consequently XAI explanations like heat maps fail to meet many users' needs. The second concerns the varying needs of interlocutors. Predictably, explanations suitable for experts and novices differ; presuppositions on explainee knowledge and goals inform explanation content. Pragmatics matter. The third difference concerns how linguistic terms and phrases are used to hedge uncertainty. There is no reason XAI might not satisfy these human requirements. To do so, however, will require deeper theories of human explanation.

Cognitive Differences in Human and AI Explanation

In some contexts, human learning greatly exceeds what the sparsity of the available data seems to allow, while in others, it can fall short, despite vast amounts of data. This apparent contradiction has led to separate explanations of humans being equipped either with background knowledge that enhances their learning or with suboptimal mechanisms that hinder it. Here, we reconcile these findings by recognising learners can be uncertain about two structural properties of environments: 1) is there only one generative model or are there multiple ones switching across time; 2) how stochastic are the generative models. We show that optimal learning under these conditions of uncertainty results in learning trade-offs: e.g., a prior for determinism fosters fast initial learning but renders learners susceptible to low asymptotic performance, when faced with high model-stochasticity. Our results reveal the existence of optimal-paths-to-not-learning and reconcile within a coherent framework, phenomena previously considered disparate.

Optimal learning under structural environmental uncertainty reveals inherent learning trade-offs

How effective do observers expect other problem solvers to be? What makes a decision seem "human"?" We addressed this question in the context of the Traveling Salesman Problem (TSP), a decision problem in which a perfectly optimal solution is intractable, but for which various kinds of approximate solutions are available. We conducted a series of experiments involving both "production tasks" in which we asked subjects to solve the TSP, and "perception tasks", in which we asked subjects to judge others' solutions to the TSP, rating them for intelligence or humanness. Results suggest that observers expect human solutions to be less optimal than algorithmic solutions: observers expect human problem solvers to exhibit a combination of local and global solution criteria, and to use a short look-ahead window when choosing a solution. These results shed light on human models of other humans' minds, a fundamental problem in social interaction and robotics.

How Optimal is Too Optimal? Expectations About Performance in the Traveling Salesman Problem

Communication involves searching for optimal utterances
within memory and then evaluating those utterances against a
target goal. This task is substantially harder when informa-
tion about multiple concepts has to be communicated, such
as describing how music and tides are similar. Whether the
search process for this challenging communicative task con-
verges onto the optimal response relatively quickly, or involves
more strategic decision-making to evaluate different candi-
dates remains understudied. In this work, speakers gener-
ated single word “clues” that would enable a listener to cor-
rectly identify a pair of words among several distractor words.
Speakers and listeners generated candidates before producing
final responses. Each player was biased towards the first can-
didate(s) they generated, even when this candidate was sub-
optimal compared to other candidates, as was the case for less
related concepts. Furthermore, straying away from the initial
semantic “patch” of responses decreased accuracy in the game.
Overall, these findings suggest that individuals tend to identify
the relevant semantic cluster early on during semantic search,
and are likely to employ the “take-the-first” strategy for select-
ing utterances in ambiguous, ill-defined semantic contexts.

Fast and frugal memory search for communication

Competing perspectives on building ethical AI: psychological, philosophical, and computational approaches

Is the mind equipped with universal biases? If so, how can we account for the well-recognized diversity ofhuman cognition? And if universal biases do exist,what is their source—do theyreflect innate principles of core knowledge, or can they emerge from domain-general pressures? Finally, how can we, cognitive scientists, “get” human nature better—minimizing the pitfalls of our intuitive cognitive biases and our narrow WEIRD perspective?This symposium seeks to address the question of human nature from a broad interdisciplinary perspective. We begin by examining the notion of “human nature” and the concept of innateness (Samuels)and exploring some of the challenges facing its discovery by the confinements of our own human cognition (Berent). The next talk (Culbertson) exposesthe tensions between cognitive universals and diversity in the domain of language, contrasts domain-specific with domain-general explanations of these facts, and demonstratesthe merits of a cross-linguistic perspective. The last talk (Majid) considershow language is used to express perceptual categories across cultures, explores how culture and nurture can shed light on human nature, and highlightssome of the challenges in sifting through the tight nature-nurture interactions. Throughout these four talks, we hope to highlight both theoretical gains and limits in our current understanding of innateness anddomain-specificity, the value of exploring human diversity, and the challenges we, humans face when we seek to explore our own human nature.

Premium content

Learning as Programming: Modeling Efficient Search in Human Concept Learning

Downloads

Next from CogSci 2022

Frequently produced semantic features reflect principled connections

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES