Canada

What information guides individuals to trust an algorithm? We examine this question across three experiments that consistently found explanations and relative performance information increased trust in an algorithm relative to a human expert. Strikingly however, in only 23% of responses (414/1800) did an individual’s preferred agent for a task (e.g., driving a car) change from human to algorithm. Thus, initial preferences were ‘sticky’ and largely resistant to large shifts in trust. We discuss theoretical and practical implications of this work and identify important contributions to our understanding of how summaries of information can improve people’s willingness to trust decision aid algorithms.

CogSci 2022

Trusting algorithms: performance, explanations, and sticky preferences

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. 





How do people build up trust with artificial agents? Here, we study a key component of interpersonal trust: people's ability to evaluate the competence of another agent across repeated interactions. Prior work has largely focused on appraisal of simple, static skills; in contrast, we probe competence evaluations in a rich setting with agents that learn over time. Participants played a video game involving physical reasoning paired with one of four artificial agents that suggested moves each round. We measure participants' decisions to accept or revise their partner's suggestions to understand how people evaluated their partner's ability. Overall, participants collaborated successfully with their agent partners; however, when revising their partner's suggestions, people made sophisticated inferences about the competence of their partner from prior behavior. Results provide a quantitative measure of how people integrate a partner's competence into their own decisions and may help facilitate better coordination between humans and artificial agents.

How do people incorporate advice from artificial agents when making physical judgments?

Deductive domains are typical of many cognitive skills in that no single problem-solving strategy is always optimal for solving all problems. It was shown that students who know how and when to use each strategy (StrTime) outperformed those who know neither and stick to the default strategy (Default). In this work, students were trained on a logic tutor that supports a default forward-chaining and a backward-chaining (BC) strategy, then a probability tutor that only supports BC. We investigated three types of interventions on teaching the Default students how and when to use which strategy on the logic tutor: Example, Nudge and Presented. Meanwhile, StrTime students received no interventions. Overall, our results show that Nudge outperformed their Default peers and caught up with StrTime on both tutors.

The Power of Nudging: Exploring Three Interventions for Metacognitive Skills Instruction across Intelligent Tutoring Systems

We extend our previous neural dynamic models of visual search and scene memory (Grieben et al. (2020); Grieben and Schöner (2021)) to move beyond classical “laboratory” stimuli. The new model can autonomously explore a natural scene and build a scene memory of recognized objects and their locations. It is also capable of guided visual search for object categories in the scene. This is achieved by learning object templates for object recognition, and feature guidance templates for visual search and associating them to categorical concepts. We address how preattentive shape can be extracted from the visual input and how scene guidance, specifically, scene gramar (Võ, 2021), emerge. For the first time, we embed feature extraction by a headless deep convolutional neural network (CNN) in a neural dynamic (DFT) architecture by learning a mapping from the distributed feature representation of the CNN to the localist representation of a dynamic neural field.

Bridging D and DNNs: A neural dynamic process model of scene representation, guided visual search and scene grammar in natural scenes

Children learn to extend novel category labels to objects with the same shape, a phenomenon known as the shape bias. Inspired by these findings, Geirhos et al. (2019) examined whether deep neural networks (DNNs) show a shape or texture bias by constructing images with conflicting shape and texture cues. They found that DNNs strongly preferred to classify familiar objects based on texture as opposed to shape. However, there are a number of differences between how the networks were tested in this study versus how children are typically tested. In this work, we re-examine the inductive biases of neural networks by adapting the stimuli and procedure from Geirhos et al. (2019) to more closely follow the developmental paradigm and test on a range of neural networks. We find that DNNs exhibit a preference for shape rather than texture when tested under conditions that more closely replicate the developmental procedure.

A Developmentally-Inspired Examination of Shape versus Texture Bias in Machines

Modern Hopfield networks (HNs) exhibit properties of a Content Addressable Memory (CAM) that can store and retrieve a large number of memories. They also provide a basis for modelling associative memory in humans. However, the implementation of these networks is often not biologically plausible as they assume the strengths of synaptic connections are symmetric, and utilize functions that rely on many-body synapses. More biologically realistic versions of Modern HNs have been proposed, although these implementations often still utilize the softmax function. Computing the softmax for a single node requires the knowledge of all other neurons, and thus still poses a degree of biological implausibility. We present a Modern HN that uses a version of softmax that can be computed in a more bio-realistic way, and hence moves us closer to a model of memory that is biologically sound. We also show that our proposed network can learn the connection weights using a local learning rule, derived from gradient descent on the energy function. Finally, we verify that our proposed biological network behaves like a Modern HN and explore some of its other interesting properties.

Biological Somax: Demonstrated in Modern Hopfield Networks

We present a new cognitive architecture that combines two neurobiologically-plausible computational elements: (1) a variant of predictive processing known as neural generative coding and (2) hyperdimensional / vector-symbolic models of human memory. We draw inspiration from well-known cognitive architectures such as ACT-R, Soar, Leabra, and Spaun/Nengo. Our cognitive architecture, the COGnitive Neural GENerative system (CogNGen), is in broad agreement with these architectures, but provides a level of detail between ACT-R’s high-level, symbolic description of human cognition and Spaun’s low-level neurobiological description. CogNGen creates the groundwork for developing agents that learn from diverse tasks and model human performance at larger scales than what is possible with existent cognitive architectures. We test CogNGen on a set of maze-learning tasks, including mazes that test short-term memory and planning, and find that the synergy between its predictive processing and vector-symbolic components allow it to master the maze tasks.

CogNGen: Building the Kernel for a Hyperdimensional Predictive Processing Cognitive Architecture

Global Workspace Theory (GWT) is a prominent account of cognitive access in humans. In the decades since its proposal, there have been a number of computational models developed to study the hypothetical dynamics of the global workspace, most of which are hand-designed to reflect the expectations of the theory. Here we examine a recently successful general deep learning architecture, the Perceiver, as a potential theoretical candidate for the global workspace. We find that despite being developed in an unrelated context, the Perceiver meets a number of theoretical requirements of the global workspace. More importantly, it demonstrates empirical behavior consistent with that expected by GWT in both attentional control and working memory tasks drawn from the cognitive science literature. Taken together, this evidence suggests that the Perceiver and related models may be a useful tool for studying the global workspace and its potential realization in both artificial and biological agents.

The Perceiver Architecture is a Functional Global Workspace

Since generations, singing and speech have been mainly transmitted orally. How does oral transmission shape the evolution of music? Here, we developed a method for conducting online transmission experiments, in which sung melodies are passed from one singer to the next. We show that cognitive  and motor constraints play a profound role in the emergence of melodic structure. Specifically, initially random tones develop into more structured systems that increasingly reuse and combine fewer elements, making melodies easier to learn and transmit over time. We discuss how our findings are compatible with melodic universals found in most human cultures and culturally specific characteristics of participants' previous musical exposure. Overall, our method efficiently automates online singing experiments while enabling large-scale data collection using standard computers available to everyone. We see great potential in further extending this work to increase the efficiency, scalability, and diversity of research on cultural evolution and cognitive science.

Studying the Effect of Oral Transmission on Melodic Structure using Online Iterated Singing Experiments

Despite the variability of music worldwide, some types of human songs share basic acoustic characteristics. For example, dance songs tend to be loud and rhythmic, whereas lullabies tend to be quiet and melodious. Prior studies with western English-speaking participants have shown that this enables listeners to infer aspects of a singer’s behavior, despite being unfamiliar with the singer’s culture and language. Here, we test whether these intuitions are shared across a diversity of languages and human societies, with 5524 people from 49 industrialised countries comprising 28 languages, and 116 people in 3 small-scale societies with limited access to global media. Each made inferences about the behavioral contexts of 118 songs from 86 societies. Both groups reliably identified the behavioral functions of dance songs, lullabies, and healing songs. Linguistic and geographical proximity between listeners and singers was minimally predictive of accuracy, demonstrating a degree of cultural invariance in music perception.

Cultural Invariance in Musical Communication

Object-centric interactions provide rich learning moments for young children, including opportunities to discover word meanings. Children’s first-person object handling experiences, in particular, form a key source of input—one that varies across cultures and across development. Using daylong photo streams from child-worn cameras, we analyze >17k images to identify the frequency and targets of child object handling across the first four years in two small-scale subsistence farming communities on opposite sides of the globe (Rossel Papuan and Tseltal Mayan). Overall, we see general consistency in the distribution of object categories (e.g., consumables, mealtime tools, natural objects, etc.) handled by children across cultures and age, likely reflecting stable properties of children’s physical environments and day-to-day routines. However, the exact objects available to children vary both within and across communities and diversify with age. These various distributions of handling patterns are discussed in their relation to potential consequences for early learning.

Trusting algorithms: performance, explanations, and sticky preferences

Downloads

Next from CogSci 2022

How do people incorporate advice from artificial agents when making physical judgments?

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES