Research Papers

We survey the treatment of sex and gender in the Computer Graphics research literature from an algorithmic fairness perspective. The established practices on the use of gender and sex in our community are scientifically incorrect and constitute a form of algorithmic bias with potential harmful effects. We propose ways of addressing these as technical limitations.

Deep learning has become a one-size-fits-all solution for technical and business domains thanks to its flexibility and adaptability. It is implemented using opaque models, which unfortunately undermines the outcome trustworthiness. In order to have a better understanding of the behavior of a system, particularly one driven by time series, a look inside a deep learning model so-called posthoc eXplainable Artificial Intelligence (XAI) approaches, is important. There are two major types of XAI for time series data, namely model-agnostic and model-specific. Model-specific approach is considered in this work. While other approaches employ either Class Activation Mapping (CAM) or Attention Mechanism, we merge the two strategies into a single system, simply called the Temporally Weighted Spatiotemporal Explainable Neural Network for Multivariate Time Series (TSEM). TSEM combines the capabilities of RNN and CNN models in such a way that RNN hidden units are employed as attention weights for the CNN feature maps temporal axis. The result shows that TSEM outperforms XCM. It is similar to STAM in terms of accuracy, while also satisfying a number of interpretability criteria, including causality, fidelity, and spatiotemporality.

The energy transition introduces more volatile energy sources into the power grids. In this context, power transfer between different synchronous areas through High Voltage Direct Current (HVDC) links becomes increasingly important. Such links can balance volatile generation by enabling long-distance transport or by leveraging their fast control behavior. Here, we investigate the interaction of power imbalances - represented through the power grid frequency - and power flows on HVDC links between synchronous areas in Europe. We use explainable machine learning to identify key dependencies and disentangle the interaction of critical features. Our results show that market-based HVDC flows introduce deterministic frequency deviations, which however can be mitigated through strict ramping limits. Moreover, varying HVDC operation modes strongly affect the interaction with the grid. In particular, we show that load-frequency control via HVDC links can both have control-like or disturbance-like impacts on frequency stability.

Predictive parity (PP), also known as sufficiency, is a core definition of algorithmic fairness essentially stating that model outputs must have the same interpretation of expected outcomes regardless of group. Testing and satisfying PP is especially important in many settings where model scores are interpreted by humans or directly provide access to opportunity, such as healthcare or banking. Solutions for PP violations have primarily been studied through the lens of model calibration. However, we find that existing calibration-based tests and mitigation methods are designed for independent data, which is often not assumable in large-scale applications such as social media or medical testing. In this work, we address this issue by developing a statistically rigorous non-parametric regression based test for PP with dependent observations. We then apply our test to illustrate that PP testing can significantly vary under the two assumptions. Lastly, we provide a mitigation solution to provide a minimally-biased post-processing transformation function to achieve PP.

Doctor XAvIer: Explainable Diagnosis on Physician-Patient Dialogues and XAI Evaluation

Hillary Ngai, Frank Rudzicz

We introduce Doctor XAvIer, a BERT-based diagnostic system that extracts relevant clinical data from transcribed patient-doctor dialogues and explains predictions using feature attribution methods. We present a novel performance plot and evaluation metric for feature attribution methods: Feature Attribution Dropping (FAD) curve and its Normalized Area Under the Curve (N-AUC). FAD curve analysis shows that integrated gradients outperforms Shapley values in explaining diagnosis classification. Doctor XAvIer outperforms the baseline with 0.97 F1-score in named entity recognition and symptom pertinence classification and 0.91 F1-score in diagnosis classification.

Training Characteristic Functions with Reinforcement Learning: XAI-methods play Connect Four

Stephan Wäldchen, Felix Huber, Sebastian Pokutta

One of the goals of Explainable AI (XAI) is to determine which input components were relevant for a classifier decision. This is commonly know as saliency attribution. Characteristic functions (from cooperative game theory) are able to evaluate partial inputs and form the basis for theoretically "fair" attribution methods like Shapley values. Given only a standard classifier function, it is unclear how partial input should be realised. Instead, most XAI-methods for black-box classifiers like neural networks consider counterfactual inputs that generally lie off-manifold. This makes them hard to evaluate and easy to manipulate. We propose a setup to directly train characteristic functions in the form of neural networks to play simple two-player games. We apply this to the game of Connect Four by randomly hiding colour information from our agents during training. This has three advantages for comparing XAI-methods: It alleviates the ambiguity about how to realise partial input, makes off-manifold evaluation unnecessary and allows us to compare the methods by letting them play against each other.

Explainability, interpretability and how much they affect human trust in AI systems are ultimately problems of human cognition as much as machine learning, yet the effectiveness of AI recommendations and the trust afforded by end-users are typically not evaluated quantitatively. We developed and validated a general purpose Human-AI interaction paradigm which quantifies the impact of AI recommendations on human decisions. In our paradigm we confronted human users with quantitative prediction tasks: asking them for a first response, before confronting them with an AI's recommendations (and explanation), and then asking the human user to provide an updated final response. The difference between final and first responses constitutes the shift or sway in the human decision which we use as metric of the AI's recommendation impact on the human, representing the trust they place on the AI. We evaluated this paradigm on hundreds of users through Amazon Mechanical Turk using a multi-branched experiment confronting users with good/poor AI systems that had good, poor or no explainability. Our proof-of-principle paradigm allows one to quantitatively compare the rapidly growing set of XAI/IAI approaches in terms of their effect on the end-user and opens up the possibility of (machine) learning trust.

Explaining Reject Options of Learning Vector Quantization Classifiers

André Artelt, Johannes Brinkrolf, Roel Visser +1

While machine learning models are usually assumed to always output a prediction, there also exist extensions in the form of reject options which allow the model to reject inputs where only a prediction with an unacceptably low certainty would be possible. With the ongoing rise of eXplainable AI, a lot of methods for explaining model predictions have been developed. However, understanding why a given input was rejected, instead of being classified by the model, is also of interest. Surprisingly, explanations of rejects have not been considered so far. We propose to use counterfactual explanations for explaining rejects and investigate how to efficiently compute counterfactual explanations of different reject options for an important class of models, namely prototype-based classifiers such as learning vector quantization models.

By defining the current limits (and thereby the frontiers), many boundaries are shaping, and will continue to shape, the future of Artificial Intelligence (AI). We push on these boundaries in order to make further progress into what were yesterday's frontiers. They are both pliable and resilient - always creating new boundaries of what AI can (or should) achieve. Among these are technical boundaries (such as processing capacity), psychological boundaries (such as human trust in AI systems), ethical boundaries (such as with AI weapons), and conceptual boundaries (such as the AI people can imagine). It is within this final category while it can play a fundamental role in all other boundaries} that we find the construct of needs and the limitations that our current concept of need places on the future AI.

An Empirical Study of Modular Bias Mitigators and Ensembles

Michael Feffer, Martin Hirzel, Samuel C. Hoffman +3

There are several bias mitigators that can reduce algorithmic bias in machine learning models but, unfortunately, the effect of mitigators on fairness is often not stable when measured across different data splits. A popular approach to train more stable models is ensemble learning. Ensembles, such as bagging, boosting, voting, or stacking, have been successful at making predictive performance more stable. One might therefore ask whether we can combine the advantages of bias mitigators and ensembles? To explore this question, we first need bias mitigators and ensembles to work together. We built an open-source library enabling the modular composition of 10 mitigators, 4 ensembles, and their corresponding hyperparameters. Based on this library, we empirically explored the space of combinations on 13 datasets, including datasets commonly used in fairness literature plus datasets newly curated by our library. Furthermore, we distilled the results into a guidance diagram for practitioners. We hope this paper will contribute towards improving stability in bias mitigation.

Language data and models demonstrate various types of bias, be it ethnic, religious, gender, or socioeconomic. AI/NLP models, when trained on the racially biased dataset, AI/NLP models instigate poor model explainability, influence user experience during decision making and thus further magnifies societal biases, raising profound ethical implications for society. The motivation of the study is to investigate how AI systems imbibe bias from data and produce unexplainable discriminatory outcomes and influence an individual's articulateness of system outcome due to the presence of racial bias features in datasets. The design of the experiment involves studying the counterfactual impact of racial bias features present in language datasets and its associated effect on the model outcome. A mixed research methodology is adopted to investigate the cross implication of biased model outcome on user experience, effect on decision-making through controlled lab experimentation. The findings provide foundation support for correlating the implication of carry-over an artificial intelligence model solving NLP task due to biased concept presented in the dataset. Further, the research outcomes justify the negative influence on users' persuasiveness that leads to alter the decision-making quotient of an individual when trying to rely on the model outcome to act. The paper bridges the gap across the harm caused in establishing poor customer trustworthiness due to an inequitable system design and provides strong support for researchers, policymakers, and data scientists to build responsible AI frameworks within organizations.

Australia is a leading AI nation with strong allies and partnerships. Australia has prioritised the development of robotics, AI, and autonomous systems to develop sovereign capability for the military. Australia commits to Article 36 reviews of all new means and methods of warfare to ensure weapons and weapons systems are operated within acceptable systems of control. Additionally, Australia has undergone significant reviews of the risks of AI to human rights and within intelligence organisations and has committed to producing ethics guidelines and frameworks in Security and Defence. Australia is committed to OECD's values-based principles for the responsible stewardship of trustworthy AI as well as adopting a set of National AI ethics principles. While Australia has not adopted an AI governance framework specifically for the Australian Defence Organisation (ADO); Defence Science and Technology Group (DSTG) has published 'A Method for Ethical AI in Defence' (MEAID) technical report which includes a framework and pragmatic tools for managing ethical and legal risks for military applications of AI. Australia can play a leadership role by integrating legal and ethical considerations into its ADO AI capability acquisition process. This requires a policy framework that defines its legal and ethical requirements, is informed by Defence industry stakeholders, and provides a practical methodology to integrate legal and ethical risk mitigation strategies into the acquisition process.