United States

Predictive coding theories argue that recent experience establishes expectations that generate prediction errors when violated. Neural correlates of prediction errors have been widely reported, but their perceptual consequences remain unknown. We had observers perform a dual-report task on the orientation of briefly presented gratings within predictable or random sequences. Observers first made a speeded response to categorize grating orientation (clockwise or counterclockwise from vertical), then reproduced the orientation without time pressure by rotating a bar. This allowed us to separately assess response speed and accuracy for the same stimulus. Critically, on half the trials, the target orientation was unexpected, due to the spatiotemporal structure established by the preceding gratings. Observers responded faster and more accurately to gratings that were unexpected. These findings reveal the behavioural consequences of prediction error and indicate that unexpected events are prioritized by the visual system both in terms of processing speed and representational fidelity.

CogSci 2023

Unexpected visual events elicit faster and more accurate behavioural responses

poster

The 45th Annual Meeting of the Cognitive Science Society will take place in Sydney, Australia as a modified hybrid conference. Presentations from the main-session room will be live-streamed every day for all events. This will include all invited talks and award sessions. Recordings of all CogSci 2023 sessions will be made available for on-demand viewing after the conference.<br>
The in-person component will be held at the International Convention Centre Sydney at Darling Harbour. The Conference will take place Wednesday July 26th – Saturday July 29th, 2023 (Wednesday, July 26th is the Workshops Day).

Join us!

<br>[**The Program at a Glance can be found here**](https://cognitivesciencesociety.org/wp-content/uploads/2023/07/COgSci-PAAG-V2.pdf)
<br>[**The Detailed Program can be found here**](https://online.fliphtml5.com/ebtyf/mswz/)
<br><br>

#### **CogSci 2023 Program Book**
<iframe style="width:700px;height:400px" src="https://online.fliphtml5.com/ebtyf/tzgk/"  seamless="seamless" scrolling="no" frameborder="0" allowtransparency="true" allowfullscreen="true" ></iframe>
<br>

The 45th Annual Meeting of the Cognitive Science Society presents the latest research across cognitive science and highlights the theme of Cognition in Context.

Seeking plausible models for brain computation has been a continuing effort in brain encoding and decoding. Most prior works have mapped the association between stimulus representation from language models and fMRI brain activity using ridge regression. However, these models are not biologically plausible from the perspective of representing neural dynamics of the brain underlying the fMRI recordings. In this work, our primary motivation is to challenge ridge regression models with simple neural architectures such as echo state network (ESNs) and long short-term memory (LSTMs) on the brain encoding task that requires full-sentence processing in the task of reading short sentences. We explore various pre-trained Transformer language models for computing sentence representations and predict the fMRI brain activity from simple neural architectures that include initial layers with random parameterization and that do not require explicit training. Experiment results show that (i) ESNs with online learning can accurately predict the fMRI brain activity comparable to ridge regression models, (ii) Both cell-state (internal memory representation related to long term memory) and out-gate (related to short term memory) of LSTM display an equal level performance during short sentences in random LSTMs, (iii) left hemisphere language area has higher predictive brain activity versus right hemisphere language area, (iv) ESNs with online learning yield superior performance over offline learning, indicating the biological plausibility of ESNs and the cognitive process of sentence reading, and (v) among all the variants of transformer models, Longformer features facilitate better accuracy when utilized with both ridge regression and ESN online learning models. The proposed framework that combines input featurization, dynamic memory and learning modules offers a flexible, biologically plausible architecture for investigating brain encoding in neuroscience.

Brain Encoding using Randomized Recurrent Networks

A common claim by advocates of embodied, dynamical approaches is that action and perception are “coupled.” On the face of it, this claim may not seem controversial, after all many researchers working in mainstream computationalist neuroscientific approaches also talk about the “coupling” of perception and action. Our goal here is to clarify the relation between these claims of perception-action coupling stemming from dynamical and from computational perspectives. Examining the empirical evidence that computationalists and dynamicists invoke to support their claims we conclude that, despite using similar terminology, they mean entirely different and incompatible things. Still, we propose that both approaches can, at least to some extent, accommodate the evidence invoked by the other. This suggests that the evidence should not, on its own, be used to argue in favor of one approach against the other, and that the disagreement is of a philosophical nature rather than an empirical one.

Perception-Action Coupling and the Dynamicist/Computationalist Divide

Effectively updating one's beliefs requires sufficient empirical evidence (i.e., data) and the computational capacity to process it. Yet both data and computational resources are limited for human minds. Here, we study the problem of belief updating under limited data and limited computation. Using information theory to characterize constraints on computation, we find that the solution to the resulting optimization problem links the data and computational limitations together: when computational resources are tight, agents may not be able to integrate new empirical evidence. The resource-rational belief updating rule we identify offers a novel interpretation of conservative Bayesian updating.

Computation-Limited Bayesian Updating

Assigning category labels to examples varying along a continuous dimension exaggerates perceived differences between members on opposite sides of category boundaries. Using social categories, we investigated how contrast may guide representation of not only features that differentiate between groups but also features that are neither diagnostic nor correlated across trained examples. In a classification task, participants learned which residence hall to correctly assign students varying along three psychological traits (academic, athletic, social). The same target category was learned alongside one of two contrast categories with either higher or lower values along one diagnostic dimension. After learning, participants provided estimates of average trait values for each dorm. Predicted contrast effects were found along diagnostic dimensions but contrast also influenced memory for non-diagnostic and uncorrelated dimensions, presumably based on assumptions about general co-occurrence of features. These findings have implications for how stereotypes are learned and applied and how illusory correlations are perpetuated.

Contrast Categories Elicit Illusory Correlations When Learning Social Groups

Categorization is ubiquitous in human cognition and society, and impacts how we perceive and understand the world. In reflecting the needs and perspectives of their creators, no categorization system is entirely objective, and inbuilt biases can have harmful social consequences. Here, we propose methods for quantifying three kinds of category biases in hierarchical category systems. We present a study on two widely used library classification systems (the DDC and LCC) as large-scale examples of human categorization, and use our methods to quantify bias towards content associated with western (vs non-western) concepts in topic areas including history and religion. We find consistent evidence for western bias and show that the DDC tends to exhibit more western bias than the LCC. Our methods are general, and can be used to survey biases across topic areas, bias attributes, and hierarchical category systems.

Quantifying Bias in Library Classification Systems

In categorization research, competing theories are typically compared by fitting their predictions to participant’s responses on a set of test items. The theory that best matches each participant's responses is identified as the strategy the participant is most likely employing. Researchers face considerable difficulty in selecting the best-fitting model due to several factors. In this study, we show the frailty of this approach. Due to pervasive model mimicry and across the similarity- and rule-based models, typical categorization task designs fail to reliably distinguish strategies. Some design modifications that might help are counter-indicated on practical grounds (e.g., carry-over effects); other possible means of improving strategy identification are also discussed.

The Categorization Task is Insufficient to Distinguish between Strategies: A Case for Partial-XOR-like Tasks

There are two hypotheses about complex pattern processing: one proposes that people are predisposed towards simple, local patterns; the other proposes that they prefer more complex hierarchical patterns. Children and adults can both learn to preferentially generate hierarchical sequences. One unanswered question is whether children and adults generate complex patterns spontaneously. We investigated the predispositions or inductive biases of human children and adults in a novel open-ended sequence generation task. We found that 3-to-6-year-old-children and adults display strong biases to generate simple patterns. Both groups overcome this inductive bias by generating more complex hierarchical patterns in subsequent learning tasks. However, children find it more difficult to override their biases. Our results suggest that people possess an inductive bias towards simple patterns that are well-organized but minimally hierarchical. With some experience, humans learn complex hierarchical patterns. These findings reveal substantial encoding flexibility for patterns in humans — a flexibility beginning in childhood.

Conceptual Change During Recursive Pattern Learning in Children and Adults

There are ongoing debates whether logarithmic compression in number-to-space mapping reflects logarithmic encoding of large numbers (bias) or uncertainty about numeric value (noise). We tested these two hypotheses by disentangling the effect of bias and noise. When 80 adults and 80 4- to 7-year-olds were asked to estimate the number of dots on a number line, both children and adults were more logarithmic on 0-100 than 0-30 problems. Internal noise explained some of the variance in logarithmicity, but only for children. We then examined the wisdom of crowds effect by comparing accuracy of children's mean estimate with accuracy of each adult's estimates. As predicted, children as a crowd were not as accurate as individual adults, indicating that noise is not the only source of children's errors. Generally, increasing the size of a crowd also had a smaller effect on 0-100 than 0-30 problems, indicating that inaccuracy on 0-30 problems is likely due to noise. The present study provides evidence that bias and noise have an additive effect on logarithmic compression and that children's logarithmicity reflects bias in number representations, not just noise.

Children Are Not Just Noisy Adults: Disentangling Noise and Bias in Numerical Estimation

We explore whether people can infer the principle of additive commutativity in the absence of cultural support—such as formal schooling and market experience. Our prior research shows that unschooled Tsimane’ adults do not succeed on abstract symbolic and word problem commutativity tasks. In a new study, we implement a concrete commutativity task with tangible tokens, including sums with specified (e.g. 2+1 vs. 1+2) and unspecified (e.g. 2+Dax vs. Dax+2) quantities. We find significantly better performance on concrete tasks than in abstract tasks, even in Tsimane’ participants with limited levels of formal schooling and other cultural supports. This suggests that cultural support is not necessary to understand the commutative principle, but is necessary to apply it to more abstract problems.

The role of cultural support on commutativity at varying levels of abstraction

People are known for their ability to learn probabilistic rewarding rules of the environment. However, the rules tested in most studies consisted of one feature dimension or a linear combination of multiple features. It is unclear how well people can learn non-linear combinatory rules where each single dimension is noninformative. In an adapted matching-pennies game, we asked participants to predict whether their opponent would choose head or tail next trial. Though the feature space for possible rules was small, including only two binary dimensions (i.e., the two players’ current choices), participants had striking failures in learning the probabilistic XOR rules (e.g., the opponent would have a probability of 0.8 to choose head if the two players had matching pennies and tail for unmatching pennies). The failure persisted in a series of experiments, even when the feature space was made explicit, after 320 trials of learning, and for different cover stories.

Next from CogSci 2023

Brain Encoding using Randomized Recurrent Networks