No abstract

This paper builds upon the theoretical foundations of the Accountable eXplainable Artificial Intelligence (AXAI) capability framework presented in part one of this paper. We demonstrate incorporation of the AXAI capability in the real time Affective State Assessment Module (ASAM) of a robotic system. We show that adhering to the eXtreme Programming (XP) practices would help in understanding user behavior and systematic incorporation of the AXAI capability in Machine Learning (ML) systems. We further show that a collaborative software design and development process (SDDP) would facilitate identification of ethical, technical, functional, and domain-specific system requirements. Meeting these requirements would increase user confidence in ML and AI systems. Our results show that the ASAM can synthesize discrete and continuous models of affective state expressions for classifying them in real-time. The ASAM continuously shares important inputs, processed data and the output information with users via a graphical user interface (GUI). Thus, the GUI presents reasons behind system decisions and disseminates information about local reasoning, data handling and decision-making. Through this demonstrated work, we expect to move toward enhancing AI systems' acceptability, utility and establishing a chain of responsibility if a system fails. We hope this work will initiate further investigations on developing the AXAI capability and use of a suitable SDDP for incorporating them in AI systems.

<p> Models of seven discrete facial expressions are built on macro-level facial muscle variations for separating distinct affective states. We propose a step-wise Hierarchical Separation and Classification Network (HSCN) that discovers dynamic and continuous macro- and micro-level variations in facial expressions. The HSCN first invokes an unsupervised cosine similarity-based separation method on continuous facial expression data and extracts twenty-one dynamic expression classes from the seven common discrete affective states. Separation between the clusters is then optimised for discovering the macro-level changes in facial muscle activations followed by splitting the upper and lower facial regions for realising and modelling changes pertaining to upper and lower facial muscle activations. A linear discriminant space is developed for clustering the upper and lower facial images on the basis of similar muscular activation patterns. Actual dynamic data and linear discriminant features are mapped for developing a rule-based expert system that would facilitate classification of twenty-one upper and twenty-one lower facial micro-expressions. Using the random forest algorithm, classification accuracies of 76.11\% were observed for dynamic macro-level facial expression classification. A support vector machine provided 73.63\% and 87.68\% accuracies respectively while classifying upper and lower facial micro-expressions. This work provides a novel framework for the dynamic assessment of affective states. Reported methods and results also provide new insight into the dynamic analysis of facial expressions of affective states. </p>

<p> Models of seven discrete facial expressions are built on macro-level facial muscle variations for separating distinct affective states. We propose a step-wise Hierarchical Separation and Classification Network (HSCN) that discovers dynamic and continuous macro- and micro-level variations in facial expressions. The HSCN first invokes an unsupervised cosine similarity-based separation method on continuous facial expression data and extracts twenty-one dynamic expression classes from the seven common discrete affective states. Separation between the clusters is then optimised for discovering the macro-level changes in facial muscle activations followed by splitting the upper and lower facial regions for realising and modelling changes pertaining to upper and lower facial muscle activations. A linear discriminant space is developed for clustering the upper and lower facial images on the basis of similar muscular activation patterns. Actual dynamic data and linear discriminant features are mapped for developing a rule-based expert system that would facilitate classification of twenty-one upper and twenty-one lower facial micro-expressions. Using the random forest algorithm, classification accuracies of 76.11\% were observed for dynamic macro-level facial expression classification. A support vector machine provided 73.63\% and 87.68\% accuracies respectively while classifying upper and lower facial micro-expressions. This work provides a novel framework for the dynamic assessment of affective states. Reported methods and results also provide new insight into the dynamic analysis of facial expressions of affective states. </p>

Like other Artificial Intelligence (AI) systems, Machine Learning (ML) applications cannot explain decisions, are marred with training-caused biases, and suffer from algorithmic limitations. Their eXplainable Artificial Intelligence (XAI) capabilities are typically measured in a two-dimensional space of explainability and accuracy ignoring the accountability aspects. During system evaluations, measures of comprehensibility, predictive accuracy and accountability remain inseparable. We propose an Accountable eXplainable Artificial Intelligence (AXAI) capability framework for facilitating separation and measurement of predictive accuracy, comprehensibility and accountability. The proposed framework, in its current form, allows assessing embedded levels of AXAI for delineating ML systems in a three-dimensional space. The AXAI framework quantifies comprehensibility in terms of the readiness of users to apply the acquired knowledge and assesses predictive accuracy in terms of the ratio of test and training data, training data size and the number of false-positive inferences. For establishing a chain of responsibility, accountability is measured in terms of the inspectability of input cues, data being processed and the output information. We demonstrate applying the framework for assessing the AXAI capabilities of three ML systems. The reported work provides bases for building AXAI capability frameworks for other genres of AI systems.