United States

Successful social coordination requires being able to predict how the other people that one depends on are likely to behave. One solution to this dilemma is to establish social conventions, which constrain individuals&#39; behavior but make prediction easier. Here, we develop a multi-agent deep reinforcement learning environment to investigate the costs associated with these conventions. In our produce-and-trade task, agents have varying production skills, but their actions must be predictable in order to be rewarded. Stronger norms improve the overall success of the group by improving the average rewards of the majority, but also systematically disadvantage agents whose specialization is in the minority of the group. Critically, this outcome is magnified by population size: as larger groups make it potentially more difficult to develop individualized representations of agents, minority agents become more likely to conform to a norm that is disadvantageous to them.


CogSci 2023

Unequal Norms Emerge Under Coordination Uncertainty in Multi-Agent Deep Reinforcement Learning

Successful social coordination requires being able to predict how the other people that one depends on are likely to behave. One solution to this dilemma is to establish social conventions, which constrain individuals' behavior but make prediction easier. Here, we develop a multi-agent deep reinforcement learning environment to investigate the costs associated with these conventions. In our produce-and-trade task, agents have varying production skills, but their actions must be predictable in order to be rewarded. Stronger norms improve the overall success of the group by improving the average rewards of the majority, but also systematically disadvantage agents whose specialization is in the minority of the group. Critically, this outcome is magnified by population size: as larger groups make it potentially more difficult to develop individualized representations of agents, minority agents become more likely to conform to a norm that is disadvantageous to them.


technical paper

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.

Humans are remarkably effective social learners, with several recent studies formalizing this capacity using computational models. 
However, previous research has often been limited to tasks where observer and demonstrator share the same reward function. In contrast, humans can learn from others who have different preferences, skills, or goals. To study social learning under individual differences, we introduce the socially correlated bandit, where participants have personalized rewards, which are correlated with but not identical to those of others. Social information can still be useful, but not when used verbatim.
We present a model of social generalization that integrates individual and social information into the generalization process, but assumes social information to be noisier and thus less informative. This model out-competes previous models, with it being the dominant strategy in evolutionary simulations.
Our findings expand on previous models of social learning, showing humans can integrate social information more flexibly than previously assumed.


Social learning with a grain of salt

Many cultures share common constellations and common narratives about the stars in the night sky. Previous research has shown that this overlap in asterisms, minimal star groupings inside constellations, is clearly present across 22 distinct culture groups and can be explained in part by properties of individual stars (brightness) and properties of pairs of stars (proximity) (Kemp, Hamacher, Little and Cropper, 2022). The same work, however, found no evidence that properties of triples (angle) and quadruples (good continuation) predicted constellation formation. We developed a behavioural experiment to explore how individuals form constellations under conditions that reduce cultural learning. We found that participants independently selected and connected similar stars, and that their responses were predicted by two properties of triples (angle and even spacing) in addition to the properties of brightness and proximity supported by previous work. Our findings lend further evidence to the theory that commonality of constellations across cultures is not a result of shared human history but rather stems from shared human nature.


Visual perception principles in constellation creation in individuals

Computational models of multiple object tracking (MOT) presuppose the existence of non-conceptual indexes in visual perception, and as a result predict that identification performance on MOT tasks should be no worse than tracking performance for the same stimuli. However, empirical evidence suggests that identification performance is worse than tracking performance in MOT. We propose a computational model of MOT that is able to account for several empirical results related to tracking performance without the use of indexes and thus avoids yoking tracking performance to identification performance. We also test our model empirically and show that it accounts well for the variation in identification performance with increasing number of targets in MOT with visually identical objects. By employing an explicit (rather than an implicit) mechanism for the maintenance of correspondence between attended locations and object identities, our model accounts for the tracking vs ID performance disparity.


Tracking Multiple Objects without Indexes

The inverse base-rate effect is a robust irrational bias that arises when people face ambiguity. The most prominent theories of this irrational bias depend on prediction error.
In this study, we gradually removed elements of a predictive learning design to test the extent to which error-driven processes underly this bias. In our first experiment, we removed explicit feedback by implementing the inverse base-rate effect in an observational learning procedure. In our second study, we further removed any causal relationship between stimulus features and category labels by moving towards an unsupervised learning procedure. This removed any information participants could use to identify category labels. In both experiments, the inverse base-rate effect persisted and remained robust. This outcome suggests that this irrational bias is independent of supervised learning procedures. We propose that any theories and models of the inverse base-rate effect must manage information encoding and connection updates without explicit prediction error. We end by proposing two clear paths for future investigations.


Errorless irrationality: removing error-driven components from the inverse base-rate effect paradigm

Recent studies suggest that errors facilitate learning in certain conditions. Despite this, reinforcement paradigms dominate learning methods, subscribing to the narrative that errorless learning is the foundation of an ideal learning environment. If we continue to view learning from this restrictive perspective, we may fail to capture and apply the benefits of errors. In this paper, we investigate two potential mechanisms of post-error learning. Participants (N = 61) learned word pairs in either a study or error trial before taking a final test. Supporting past error learning literature, errors before a study opportunity led to better performance on a final test. Differences in reaction times between conditions support the theory that errors increase learning by acting as a mediator, or secondary cue, to the correct answer on subsequent tests.


Faulty Memories, Favored Outcomes: How Errors Impact Learning Processes

Efforts to correct misconceptions have produces multiple types of complex interventions. However, past research has consistently shown that simple, factual statements may be equally effective in activating the incorrect information and prompting an integration of the correct concept. The current study tests the effectiveness of simple statements on 76 different misconceptions. Findings indicate that for the majority of misconceptions, simple statements are an effective intervention to reduce misconceptions. Limitations and future directions are discussed.


Just Tell the Truth: Correcting Misconceptions with Simple, Factual Statements

Psycholinguistic analyses provide a means of evaluating large language model (LLM) output and making systematic comparisons to human-generated text. These methods can be used to characterize the psycholinguistic properties of LLM output and illustrate areas where LLMs fall short in comparison to human-generated text. In this work, we apply psycholinguistic methods to evaluate the psycholinguistic properties within sentences from student-generated long-form analogies about biochemical concepts. We compare analogies generated by human subjects enrolled in introductory biochemistry courses to analogies generated by chatGPT. We use 78 features extracted from Coh-metrix that analyze text cohesion, language, and readability (Graesser, McNamara, Louwerse, & Cai, 2004). We perform an unsupervised cluster analysis. Results illustrate a high level of performance for clustering student-generated and chatGPT-generated analogies. To evaluate which features contribute most to model performance, we use a hierarchical cluster analysis. Results from the hierarchical cluster analysis illustrate several linguistic differences between the two groups.


Long-form analogies generated by chatGPT lack human-like psycholinguistic properties

Regular polysemes are sets of ambiguous words that all
share the same relationship between their meanings, such
as CHICKEN and LOBSTER both referring to an an-
imal or its meat. To probe how a context embedding
model, here exemplified by BERT, represents regular pol-
ysemy, we analyzed whether its embeddings support an-
swering sense analogy questions similar to “is the map-
ping between CHICKEN (as an animal) to CHICKEN (as
a meat) the same as that which maps between LOBSTER
(as an animal) and LOBSTER (as a meat)?” We found
that (1) the model was sensitive to the shared structure
within a regularity type; (2) the shared structure varies
across regularity types, potentially reflective of a “reg-
ularity continuum;” (3) some high-order latent structure
may be shared across regularity types, suggestive of a
similar latent structure across types; and (4) the is equivo-
cal evidence that the aforementioned effects are explained
by meaning overlap.


Premium content

Unequal Norms Emerge Under Coordination Uncertainty in Multi-Agent Deep Reinforcement Learning

Downloads

Next from CogSci 2023

Social learning with a grain of salt

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES