skip to main content

Attention:

The NSF Public Access Repository (NSF-PAR) system and access will be unavailable from 5:00 PM ET until 11:00 PM ET on Friday, June 21 due to maintenance. We apologize for the inconvenience.


Title: Abstraction, validation , and generalization for explainable artificial intelligence
Abstract

Neural network architectures are achieving superhuman performance on an expanding range of tasks. To effectively and safely deploy these systems, their decision‐making must be understandable to a wide range of stakeholders. Methods to explain artificial intelligence (AI) have been proposed to answer this challenge, but a lack of theory impedes the development of systematic abstractions, which are necessary for cumulative knowledge gains. We propose Bayesian Teaching as a framework for unifying explainable AI (XAI) by integrating machine learning and human learning. Bayesian Teaching formalizes explanation as a communication act of an explainer to shift the beliefs of an explainee. This formalization decomposes a wide range of XAI methods into four components: (a) the target inference, (b) the explanation, (c) the explainee model, and (d) the explainer model. The abstraction afforded by Bayesian Teaching to decompose XAI methods elucidates the invariances among them. The decomposition of XAI systems enables modular validation, as each of the first three components listed can be tested semi‐independently. This decomposition also promotes generalization through recombination of components from different XAI systems, which facilitates the generation of novel variants. These new variants need not be evaluated one by one provided that each component has been validated, leading to an exponential decrease in development time. Finally, by making the goal of explanation explicit, Bayesian Teaching helps developers to assess how suitable an XAI system is for its intended real‐world use case. Thus, Bayesian Teaching provides a theoretical framework that encourages systematic, scientific investigation of XAI.

 
more » « less
Award ID(s):
1828528
NSF-PAR ID:
10361732
Author(s) / Creator(s):
 ;  ;  
Publisher / Repository:
Wiley Blackwell (John Wiley & Sons)
Date Published:
Journal Name:
Applied AI Letters
Volume:
2
Issue:
4
ISSN:
2689-5595
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. null (Ed.)
    Abstract State-of-the-art deep-learning systems use decision rules that are challenging for humans to model. Explainable AI (XAI) attempts to improve human understanding but rarely accounts for how people typically reason about unfamiliar agents. We propose explicitly modelling the human explainee via Bayesian teaching, which evaluates explanations by how much they shift explainees’ inferences toward a desired goal. We assess Bayesian teaching in a binary image classification task across a variety of contexts. Absent intervention, participants predict that the AI’s classifications will match their own, but explanations generated by Bayesian teaching improve their ability to predict the AI’s judgements by moving them away from this prior belief. Bayesian teaching further allows each case to be broken down into sub-examples (here saliency maps). These sub-examples complement whole examples by improving error detection for familiar categories, whereas whole examples help predict correct AI judgements of unfamiliar cases. 
    more » « less
  2. Recent development in the field of explainable artificial intelligence (XAI) has helped improve trust in Machine-Learning-as-a-Service (MLaaS) systems, in which an explanation is provided together with the model prediction in response to each query. However, XAI also opens a door for adversaries to gain insights into the black-box models in MLaaS, thereby making the models more vulnerable to several attacks. For example, feature-based explanations (e.g., SHAP) could expose the top important features that a black-box model focuses on. Such disclosure has been exploited to craft effective backdoor triggers against malware classifiers. To address this trade-off, we introduce a new concept of achieving local differential privacy (LDP) in the explanations, and from that we establish a defense, called XRand, against such attacks. We show that our mechanism restricts the information that the adversary can learn about the top important features, while maintaining the faithfulness of the explanations. 
    more » « less
  3. Objective Over the past decade, we developed and studied a face-to-face video-based analysis-of-practice professional development (PD) model. In a cluster randomized trial, we found that the face-to-face model enhanced elementary science teacher knowledge and practice and resulted in important improvements to student science achievement (student treatment effect, d = 0.52; Taylor et al, 2017; Roth et al, 2018). The face-to-face PD model is expensive and difficult to scale. In this paper, we present the results of a two-year design-based research study to translate the face-to-face PD into a facilitated online PD experience. The purpose is to create an effective, flexible, and cost-efficient PD model that will reach a broader audience of teachers. Perspective/Theoretical Framework The face-to-face PD model is grounded in situated cognition and cognitive apprenticeship frameworks. Teachers engage in learning science content and effective science teaching practices in the context in which they will be teaching. There are scaffolded opportunities for teachers to learn from analysis of model videos by experienced teachers, to try teaching model units, to analyze video of their own teaching efforts, and ultimately to develop their own unit, with guidance. The PD model attends to the key features of effective PD as described by Desimone (2009) and others. We adhered closely to the design principles of the face-to-face model as described by Authors, 2019. Methods We followed a design-based research approach (DBR; Cobb et al., 2003; Shavelson et al., 2003) to examine the online program components and how they promoted or interfered with the development of teachers’ knowledge and reflective practice. Of central interest was the examination of mechanisms for facilitating teacher learning (Confrey, 2006). To accomplish this goal, design researchers engaged in iterative cycles of problem analysis, design, implementation, examination, and redesign (Wang & Hannafin, 2005) in phase one of the project before studying its effect. Data Three small pilot groups of teachers engaged in both synchronous and asynchronous components of the larger online course which began implementation with a 10-week summer course that leads into study groups of participants meeting through one academic year. We iteratively designed, tested, and revised 17 modules across three pilot versions. On average, pilot groups completed one module every two weeks. Pilot 1 began the work in May 2019; Pilot 2 began in August 2019, and Pilot 3 began in October 2019. Pilot teachers responded to surveys and took part in interviews related to the PD. The PD facilitators took extensive notes after each iteration. The development team met weekly to discuss revisions. We revised all modules between each pilot group and used what we learned to inform our development of later modules within each pilot. For example, we applied what we learned from testing Module 3 with Pilot 1 to the development of Module 3 for Pilots 2, and also applied what we learned from Module 3 with Pilot 1 to the development of Module 7 for Pilot 1. Results We found that community building required the same incremental trust-building activities that occur in face-to-face PD. Teachers began with low-risk activities and gradually engaged in activities that required greater vulnerability (sharing a video of themselves teaching a model unit for analysis and critique by the group). We also identified how to contextualize technical tools with instructional prompts to allow teachers to productively interact with one another about science ideas asynchronously. As part of that effort, we crafted crux questions to surface teachers’ confusions or challenges related to content or pedagogy. We called them crux questions because they revealed teachers’ uncertainty and deepened learning during the discussion. Facilitators leveraged asynchronous responses to crux questions in the synchronous sessions to push teacher thinking further than would have otherwise been possible in a 2-hour synchronous video-conference. Significance Supporting teachers with effective, flexible, and cost-efficient PD is difficult under the best of circumstances. In the era of covid-19, online PD has taken on new urgency. NARST members will gain insight into the translation of an effective face-to-face PD model to an online environment. 
    more » « less
  4. Recent years have witnessed the growing literature in empirical evaluation of explainable AI (XAI) methods. This study contributes to this ongoing conversation by presenting a comparison on the effects of a set of established XAI methods in AI-assisted decision making. Based on our review of previous literature, we highlight three desirable properties that ideal AI explanations should satisfy — improve people’s understanding of the AI model, help people recognize the model uncertainty, and support people’s calibrated trust in the model. Through three randomized controlled experiments, we evaluate whether four types of common model-agnostic explainable AI methods satisfy these properties on two types of AI models of varying levels of complexity, and in two kinds of decision making contexts where people perceive themselves as having different levels of domain expertise. Our results demonstrate that many AI explanations do not satisfy any of the desirable properties when used on decision making tasks that people have little domain expertise in. On decision making tasks that people are more knowledgeable, the feature contribution explanation is shown to satisfy more desiderata of AI explanations, even when the AI model is inherently complex. We conclude by discussing the implications of our study for improving the design of XAI methods to better support human decision making, and for advancing more rigorous empirical evaluation of XAI methods. 
    more » « less
  5. Abstract Background

    Providing adaptive scaffolds to help learners develop effective self‐regulated learning (SRL) behaviours has been an important goal for intelligent learning environments. Adaptive scaffolding is especially important in open‐ended learning environments (OELE), where novice learners often face difficulties in completing their learning tasks.

    Objectives

    This paper presents a systematic framework for adaptive scaffolding in Betty's Brain, a learning‐by‐teaching OELE for middle school science, where students construct a causal model to teach a virtual agent, generically named Betty. We evaluate the adaptive scaffolding framework and discuss its implications on the development of more effective scaffolds for SRL in OELEs.

    Methods

    We detect key cognitive/metacognitiveinflection points, that is, moments where students' behaviours and performance change during learning, often suggesting an inability to apply effective learning strategies. At inflection points, Mr. Davis (a mentor agent in Betty's Brain) or Betty (the teachable agent) provides context‐specific conversational feedback, focusing on strategies to help the student become a more productive learner, or encouragement to support positive emotions. We conduct a classroom study with 98 middle schoolers to analyse the impact of adaptive scaffolds on students' learning behaviours and performance. We analyse how students with differential pre‐to‐post learning outcomes receive and use the scaffolds to support their subsequent learning process in Betty's Brain.

    Results and Conclusions

    Adaptive scaffolding produced mixed results, with some scaffolds (viz., strategic hints that supported debugging and assessment of causal models) being generally more useful to students than others (viz., encouragement prompts). Additionally, there were differences in how students with high versus low learning outcomes responded to some hints, as suggested by the differences in their learning behaviours and performance in the intervals after scaffolding. Overall, our findings suggest how adaptive scaffolding in OELEs like Betty's Brain can be further improved to better support SRL behaviours and narrow the learning outcomes gap between high and low performing students.

    Implications

    This paper contributes to our understanding and impact of adaptive scaffolding in OELEs. The results of our study indicate that successful scaffolding has to combine context‐sensitive inflection points with conversational feedback that is tailored to the students' current proficiency levels and needs. Also, our conceptual framework can be used to design adaptive scaffolds that help students develop and apply SRL behaviours in other computer‐based learning environments.

     
    more » « less