Canada

The introduction of deep learning and CNNs to image recognition problems has led to state-of-the-art classification accuracy. However, CNNs exacerbate the issue of algorithm explainability due to deep learning’s black box nature. Numerous explainable AI (XAI) algorithms have been developed that provide developers insight into the operations of deep learning. We aim to make XAI explanations more user-centric by introducing modifications to existing XAI algorithms based on cognitive theory. The goal of this research is to yield intuitive XAI explanations that more closely resemble explanations given by experts in the domain of bird watching. Using an existing base XAI algorithm, we conducted two user studies with expert bird watchers and found that our novel averaged and contrasting XAI algorithms are significantly preferred over the base XAI algorithm for bird identification.

CogSci 2022

User-Centric Enhancements to Explainable AI Algorithms for Image Classification

poster

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. 





We demonstrate how medical-image classification decisions can be denoised by aggregating decisions on similar images. In our algorithm, the final decision on a target image is cancerous if a percentage t of the k most similar images are cancerous, else it is not cancerous. Similarity between images is calculated as the distance between representations from an artificial neural network. We vary k and t for novice and expert participants using data from Trueblood et al. (2018) and Trueblood et al. (2021). We show that increasing k improves performance for novices, with their performance approaching that of experts. We also show that the algorithm is biased towards identifying cancerous cells, which is reflected in the representational space. The percentage t allows greater control over sensitivity and specificity and can be used to debias decisions. This algorithm is less effective for experts, partially explained by them giving similar responses on similar images.

Representational Smoothing to Improve Medical Image Decision Making

Cognitive theories for reasoning are about understanding how humans come to conclusions from a set of premises. Starting from hypothetical thoughts, we are interested which are the implications behind basic everyday language and how do we reason with them. A widely studied topic is whether cognitive theories can account for typical reasoning tasks and be confirmed by own empirical experiments. This paper takes a different view and we do not propose a theory, but instead take findings from the literature and show how these, formalized as cognitive principles within a logical framework, can establish a quantitative notion of reasoning, which we call plausibility. For this purpose, we employ techniques from non-monotonic reasoning and computer science, namely, a solving paradigm called answer set programming (ASP). Finally, we can fruitfully use plausibility reasoning in ASP to test the effects of an existing experiment and explain different majority responses.

A Quantitative Symbolic Approach to Individual Human Reasoning

A central component of the predictive coding theoretical framework concerns the comparison between predictions and sensory decoding. In the probabilistic setting, this takes the form of assessing the similarity or distance between probability distributions. However, such similarity or distance measures are not associated with explicit probabilistic models, making their assumptions implicit. In this paper, we explore an original variation on probabilistic coherence variables; we define a probabilistic component, that we call a "Bayesian comparator", that mathematically yields a particular similarity measure. A geometrical analogy suggests two variants of this measure. We apply these similarity measures to simulate the comparison of known, predicted patterns to patterns from sensory decoding, first in a simple, illustrative model, and second, in a previous model of visual word recognition. Experimental results suggest that the variant that is scaled by the norms of both predicted and perceived probability distributions yields better robustness and more desirable dynamics.

Bayesian comparators: a probabilistic modeling tool for similarity evaluation between predicted and perceived patterns

The ability to infer general characteristics of populations from specific instances is critical for reasoning. While there is evidence of this capacity in infancy, prior work has not examined children’s ability to use these second-order inferences to make predictions about future outcomes. In the current study, 3-year-olds observed balls drawn at random from two containers. In one sample each ball was a different color. The other sample consisted of balls of only one (Experiment 1) or two (Experiment 2) colors. Children were asked which of the containers was more likely to contain a novel colored ball. A significant majority of children chose the more variable sample’s container. This suggests that 3-year-olds are not only able to make inferences about hidden populations from the variability of observed samples, but also use those inferences to reason beyond their direct experience.

Reasoning from Samples to Populations: Children Use Variability Information to Predict Novel Outcomes

The "myside bias'' in evaluating arguments is an empirically well-confirmed phenomenon that consists of overweighting arguments that endorse one's beliefs or attack alternative beliefs while underweighting arguments that attack one's beliefs or defend alternative beliefs. This paper makes two contributions: First, it proposes a probabilistic model that adequately captures three salient features of myside bias in argument evaluation. Second, it provides a Bayesian justification of this model, thus showing that myside bias has a rational Bayesian explanation under certain conditions.

The Myside Bias in Argument Evaluation: A Bayesian Model

Previous work has demonstrated that relational responding in the Relational Match-to-Sample (RMTS) task can be improved compared to baseline by providing people with the opportunity to consider individual problem elements in isolation prior to receiving the full problem. However, it remains unclear whether the benefits of this intervention persist in its absence, or can be observed when a prior strategy has already been established to complete the task. To this end, we deployed a 2x2 mixed design, where presentation type was the within-subjects factor and order of presentation type was the between-subjects factor. The aims of this work were to gain further insight into the utility of isolated-focus as a means of promoting relational responding, and better understanding how processing of problem elements changes as a consequence of isolated-focus. We discuss the implications for pedagogical practices and extensions of this work to other materials and tasks.

Examining Strategy Differences on the Relational Match-to-Sample Task (RMTS)

Several individual differences have been suggested to explain neutral or negative effects of feedback on learning (Hattie & Timperley, 2007; Kluger & DeNisi, 1996). Specifically, Fyfe and colleagues (2016) have found that feedback has been useful for students with low prior knowledge, but has mixed effects on students with high prior knowledge. In contrast to prior studies that measure prior knowledge, here we manipulate students' propensity to transfer from knowledge of a more familiar concept to a more difficult one using knowledge activation (Sidney, 2020). Undergraduates (N = 138) were randomly assigned in a 3 (delayed, immediate, or no feedback) x 2 (activate prior knowledge or not) between-subjects design. Both activating prior knowledge and providing immediate feedback enhanced performance during learning trials. On a no-feedback posttest, feedback effects were moderated by knowledge activation. Importantly, combining knowledge activation with immediate feedback had a large positive impact on performance.

Role of Prior Knowledge in Feedback Timing

Variation in second language acquisition is evident from earliest stages. This study examined effects of learning tasks (retrieval practice, comprehension, verbal repetition) on comprehension of Turkish as a new language. Undergraduates (N = 156) engaged with Turkish spoken dialogues in a computer-assisted language learning session via Zoom, with learning tasks manipulated between-subjects. Participants completed pre/posttests assessing comprehension of Turkish number and case marking, a vocabulary test, and open-response questions gauging explicit awareness. The retrieval-practice group showed highest performance overall, after controlling for significant effects of nonverbal ability and pretest. For comprehension of number/case marking, the comprehension group performed comparably to the retrieval-practice group. For vocabulary comprehension, the verbal-repetition group performed comparably to the retrieval-practice group. Differential performance associated with learning tasks indicates benefits of testing and production and aligns with transfer-appropriate processing. As predicted by the noticing hypothesis, explicit awareness of number and case marking correlated with comprehension accuracy.

Retrieval Practice Promotes Learning of Turkish as a Foreign Language: A Computer-Assisted Language Learning Study

A notional machine is an abstract representation of how a computer works. Prior research shows that students often form incomplete or inaccurate mental models of the notional machine, reflecting misunderstandings and causing programming errors. To date, however, there is very little work on how to help students acquire accurate models of the notional machine. To fill this gap, this research focuses on novice programmers and experimentally investigates the impact of direct instruction of the notional machine compared to standard instruction that does not include the notional machine. The instruction is embedded in an online computer tutor that we implemented and compared to a version of the tutor that does not refer to the notional machine but otherwise includes the same materials. In particular, both tutor versions include instructional videos followed by self-explanation prompts. We will present data from an evaluation of both tutor versions.

Self-Explaining the Notional Machine to Improve Novice Programmers’ Learning and Mental Models, Supported by a Computer Tutor System

The dual mechanisms of control framework describes two modes of goal-directed behavior: proactive control (goal maintenance) and reactive control (goal activation on task demands). Shifts between these modes may explain variations in user performance in computing tasks. Although these mechanisms are relevant to the design of interactive systems, and particularly educational systems, their relation to human-computer interaction (HCI) is under-researched. We propose a manipulation to induce proactive or reactive control in the context of mathematical problem solving on an online tutoring system. We present two experiments where students solved problems using either proactive or reactive control. Study 1 validates the manipulation by investigating behavioral measures that reflect usage of the intended strategy and assesses whether either cognitive control mode impacted learning. Study 2 investigates whether alternating between control modes during problem solving affects student performance.

Premium content

User-Centric Enhancements to Explainable AI Algorithms for Image Classification

Downloads

Next from CogSci 2022

Representational Smoothing to Improve Medical Image Decision Making

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES