United States

People learn by both decomposing and combining concepts; most accounts of combination are either compositional or conjunctive. We augment the DORA model of representation learning to build new predicate representation by combining (or compressing) existing predicate representations (e.g., building a predicate a_b by combining predicates a and b). The resulting model learns structured relational representations from experience and then combines these relational concepts to form more complex, compressed concepts. We show that the resulting model provides an account of a category learning experiment in which categories are defined as novel combinations of relational concepts.

CogSci 2021

Compression: A Lossless Mechanism for Learning Complex Structured Relational Representations

### Looking for Cogsci 2023? Please [click here](https://underline.io/events/405/reception) to switch to 2023 edition.


**Dear Colleagues,**

Welcome to the 43rd Annual Conference of the Cognitive Science Society! 

This year’s conference brings together the latest research from all Cognitive Science and also highlights this year’s theme **Comparative Cognition: Animal Minds.** 

The program features four plenary speakers: Friederike Range (University of Veterinary Medicine, Vienna), Onur Güntürkün (Ruhr University Bochum), Julia Fischer (University of Göttingen), and Erich Jarvis (The Rockefeller University). The keynotes and invited symposia showcase the broad spectrum of ideas that encompass the conference theme of comparative cognition. 

Further, the program includes the Rumelhart Prize keynote presentation by Susan Goldin-Meadow (University of Chicago), and symposia to celebrate the winner of the Elman Prize: Wei Ji Ma (New York University), the Heineken Award winner Robert Zatorre (McGill University), and the Robert J. Glushko Dissertation Prizes session.

The program committee for CogSci 2021 received 1,025 submissions, including 710 full papers, 275 member abstracts, 8 publication-based talks, as well as 15 proposals for symposia, and 17 for workshops and tutorials. After a rigorous review process, the committee selected 152 papers for oral presentation and inclusion in the conference proceedings (14.8%), 332 papers for poster presentation and inclusion in the proceedings (32.4%), and 175 papers for poster presentation with inclusion of abstracts in the proceedings (17.1%) We also selected 275 submitted member abstracts. In addition, we accepted 6 publication-based talks, 12 symposia, and 9 workshops and tutorials to make for an exciting and inclusive program.

We hope that you enjoy the conference this year!<br><br>

**Your Program Chairs,**
![](https://assets.underline.io/uploads/markdown_image/1/image/99984908caba5e6453e5fc5ba34f3507.jpg)

**Learn more about the Vienna Cognitive Science Hub at the University of Vienna, hosts of CogSci 2021**

<iframe width="560" height="315" src="https://www.youtube.com/embed/464h1m3iU68" title="YouTube video player" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture" allowfullscreen></iframe>

To review the proceedings for the conference or review the searchable program, please visit the [website](https://cognitivesciencesociety.org/program/).


The Cognitive Science Society brings together researchers from around the world who hold a common goal: understanding the nature of the human mind. The mission of the Society is to promote Cognitive Science as a discipline, and to foster scientific interchange among researchers in various areas of study, including Artificial Intelligence, Linguistics, Anthropology, Psychology, Neuroscience, Philosophy, and Education. The Society is a non-profit professional organization and its activities include sponsoring an annual conference and publishing the journals Cognitive Science and TopiCS.

technical paper

Categorizing spatial relations is central to the development of visual understanding and spatial cognition, with roots in the first few months of life. Quinn (2003) reviews two findings in infant relation categorization: categorizing one object as above/below another precedes categorizing an object as between other objects, and categorizing relations over specific objects predates abstract relations over varying objects. We model these phenomena with deep neural networks, including contemporary architectures specialized for relational learning and vision models pretrained on baby headcam footage (Sullivan et al., 2020). Across two computational experiments, we can account for most of the developmental findings, suggesting these neural network models are useful for studying the computational mechanisms of infant categorization.

Examining Infant Relation Categorization Through Deep Neural Networks

Improving the accuracy of medical image interpretation is critical to improving the diagnosis of many diseases. Using both novices (undergraduates) and experts (medical professionals), we investigate methods for improving the accuracy of a single decision maker by aggregating repeated decisions from an individual in different ways. Our participants made classification decisions (cancerous versus non-cancerous) and confidence judgments on a series of cell images, viewing and classifying each image twice. We first applied the maximum confidence slating algorithm (Koriat, 2012), which leverages metacognitive ability by using the most confident response for an image as the `final response'. We also examined algorithms that aggregated decisions based on image similarity, leveraging neural network models to determine similarity. We found maximum confidence slating improves classification accuracy for both novices and experts. However, aggregating responses on similar images improves classification accuracy for novices and not experts, suggesting differences in the decision mechanisms of novices and experts.

Improving Medical Image Decision Making by Leveraging Metacognitive Processes and Representational Similarity

Recent work (Xu et al., 2020) has suggested that numeral systems  in  different  languages  are  shaped  by  a  functional  need for efficient communication in an information-theoretic sense. Here we take a learning-theoretic approach and show how efficient communication emerges via reinforcement learning.  In our  framework,  two  artificial  agents  play  a  Lewis  signaling game  where  the  goal  is  to  convey  a  numeral  concept.   The agents  gradually  learn  to  communicate  using  reinforcement learning and the resulting numeral systems are shown to be efficient in the information-theoretic framework of Regier et al.(2015); Gibson et al. (2017).  They are also shown to be similar to human numeral systems of same type.  Our results thus provide a mechanistic explanation via reinforcement learning of the recent results in Xu et al. (2020) and can potentially be generalized to other semantic domains.

Learning Approximate and Exact Numeral Systems via Reinforcement Learning

Estimating parameters of cognitive models is crucial to be able to accurately describe cognitive processing of individuals, under varying circumstances. To ensure that individual parameter estimates represent individual cognitive processes, it is important to consider model identification and specific influence. Model identification refers to whether a unique set of parameter estimates is associated with a particular pattern in the data; Specific influence means that certain experimental manipulations affect only specific cognitive processes, reflected in changes in only those parameters that represent those processes, and not others. These two general concepts also apply to cognitive models of more applied tasks and settings, such as driving. In the current work, we specifically test whether these two requirements hold in a commonly used cognitive model of visual control of steering behavior. For this model, we test the identifiability, and then estimate parameters of two experiments to understand how cognitive load and driving speed specifically influence parameter estimates of the model. The results indicate that the two-point visual control of steering model is identified, and that cognitive load and driving speed are related to different parameters.

Identifiability and Specificity of the Two-Point Visual Control Model of Steering

Curiosity is a fundamental driver of human behavior, and yet because of its open-ended nature and the wide variety of behaviors it inspires in different contexts, it is remarkably difficult to study in a laboratory context. A promising approach to developing and testing theories of curiosity is to instantiate them in artificial agents that are able to act and explore in a simulated environment, and then compare the behavior of these agents to humans exploring the same stimuli. Here we propose a new experimental paradigm for examining children's -- and AI agents' -- curiosity about objects' physical interactions. We let them choose which object to drop another object onto in order to create the most interesting effect. We compared adults' (N=155) and children's choices (N=66; 3-7 year-olds) and found that both children and adults show a strong preference for choosing target objects that could potentially contain the dropped object. Adults alone also make choices consistent with achieving support relations. We contextualize our results using heuristic computational models based on 3D physical simulations of the same scenarios judged by participants.

Predicting children's and adults' preferences in physical interactions via physics simulation

What do infants and young children tend to see in their everyday lives? Relatively little work has examined the categories and objects that tend to be in the infant view during everyday experience, despite the fact that this knowledge is central to theories of category learning. Here, we analyzed the prevalence of the categories (e.g., people, animals, food) in the infant view in a longitudinal dataset of egocentric infant visual experience. Overall, we found a surprising amount of consistency in the broad characteristics of children's visual environment across individuals and across developmental time, in contrast to prior work examining the changing nature of the social signals in the infant view. In addition, we analyzed the distribution and identity of the categories that children tended touch and interact with in this dataset, generalizing previous findings that these objects tended to be distributed in a Zipfian manner. Taken together, these findings take a first step towards characterizing infants' changing visual environment, and call for future work to examine the generalizability of these results and to link them to learning outcomes.

Characterizing the object categories two children see and interact with in a dense dataset of naturalistic visual experience

Infants learn the meaning of words from accumulated experiences of real-time interactions with their caregivers. To study the effects of visual sensory input on word learning, we recorded infant's view of the world using head-mounted eye trackers during free-flowing play with a caregiver. While playing, infants were exposed to novel label-object mappings and later learning outcomes for these items were tested after the play session. In this study we use a classification based approach to link properties of infants' visual scenes during naturalistic labeling moments to their word learning outcomes. We find that a model which integrates both highly informative and ambiguous sensory evidence is a better fit to infants' individual learning outcomes than models where either type of evidence is taken alone, and that raw labeling frequency is unable to account for the word learning differences we observe. Here we demonstrate how a computational model, using only raw pixels taken from the egocentric scene image, can derive insights on human language learning.

In-the-Moment Visual Information from the Infant's Egocentric View Determines the Success of Infant Word Learning: A Computational Study

Collecting new information about the outside world is a key aspect of brain function. In the context of vision, we move our eyes multiple times per second to accumulate evidence about a scene. Prior studies have suggested that this process is goal-directed and close to optimal. Here, we show that this process of seeking new information suffers from a confirmation bias similar to what has been observed in a wide range of other contexts. We present data from a new gaze-contingent task that allows us to both estimate a participant's current belief, and compare that to their subsequent eye-movements. We find that these eye-movements are biased in a confirmatory way. Finally, we show that these empirical results can be parsimoniously explained under the assumption that the brain performs approximate, not exact, inference, with computations being more approximate in decision-making compared to sensory areas.

A confirmation bias due to approximate active inference

Decision-making is often hierarchical and approximate in nature: decisions are not being made based on actual observations, but on intermediate variables that themselves have to be inferred. Recently, we showed that during sequential perceptual decision-making, those conditions induce characteristic temporal biases that depend on the balance of sensory and category information present in the stimulus. Here, we show that the same model makes predictions for when observers will be over-confident and when they will be under-confident with respect to a Bayesian observer. We tested these predictions by collecting new data in a dual-report decision-making task. We found that for most participants the bias in confidence judgments changed in the predicted direction for stimulus changes that led them from over-weighting early evidence to equal weighting of evidence or over-weighting of late evidence. Our results suggest that approximate hierarchical inference might provide the computational basis for biases beyond low-level perceptual decision-making, including those affecting higher level cognitive functions like confidence judgements.

Relating confidence judgements to temporal biases in perceptual decision-making

Two very different kinds of views exist of how infants learn about physics: on the one hand, through domain-specific rules, on the other, through general purpose simulation.
We attempt to reconcile these two views through a model that uses simulation to bootstrap rule learning. This model makes a variety of predictions about rapid concept acquisition in young infants which are consistent with experiments performed by developmental cognitive scientists. Consistent with the developmental literature, our model learns physical rules from just a few examples, but only when those examples are consistent with general physical principles. A model without simulation shows no such biases. Our approach provides a general mechanism for explaining how simulation and rule learning might bootstrap off of each other throughout development, and opens up a number of new questions about how children learn physical representations.

Downloads

Next from CogSci 2021

Examining Infant Relation Categorization Through Deep Neural Networks

Similar lecture

Students Prefer to Learn from Figures that Include Spatial Supports for Comparison

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from CogSci 2021

Examining Infant Relation Categorization Through Deep Neural Networks

Similar lecture

Students Prefer to Learn from Figures that Include Spatial Supports for Comparison

Downloads