DARPA formulated the Explainable Artificial Intelligence (XAI) program in 2015 with the goal to enable end users to better understand, trust, and effectively manage artificially intelligent systems. In 2017, the four-year XAI research program began. Now, as XAI comes to an end in 2021, it is time to reflect on what succeeded, what failed, and what was learned. This article summarizes the goals, organization, and research progress of the XAI Program.

More accurate predictions of the biological properties of chemical compounds would guide the selection and design of new compounds in drug discovery and help to address the enormous cost and low success-rate of pharmaceutical R&D. However this domain presents a significant challenge for AI methods due to the sparsity of compound data and the noise inherent in results from biological experiments. In this paper, we demonstrate how data imputation using deep learning provides substantial improvements over quantitative structure-activity relationship (QSAR) machine learning models that are widely applied in drug discovery. We present the largest-to-date successful application of deep-learning imputation to datasets which are comparable in size to the corporate data repository of a pharmaceutical company (678,994 compounds by 1166 endpoints). We demonstrate this improvement for three areas of practical application linked to distinct use cases; i) target activity data compiled from a range of drug discovery projects, ii) a high value and heterogeneous dataset covering complex absorption, distribution, metabolism and elimination properties and, iii) high throughput screening data, testing the algorithm’s limits on early-stage noisy and very sparse data. Achieving median coefficients of determination, R2, of 0.69, 0.36 and 0.43 respectively across these applications, the deep learning imputation method offers an unambiguous improvement over random forest QSAR methods, which achieve median R2 values of 0.28, 0.19 and 0.23 respectively. We also demonstrate that robust estimates of the uncertainties in the predicted values correlate strongly with the accuracies in prediction, enabling greater confidence in decision-making based on the imputed values.

COGLE (COmmon Ground Learning and Explanation) is an explainable artificial intelligence (XAI) system for autonomous drones that deliver supplies in mountainous areas to field units. The drone missions have risks that vary with topography, flight decisions, and mission goals in a simulated environment. Users must determine which AI-controlled drone is better for a mission. Narrative explanations identify the advantages of a drone’s plan (“What?”) and reasons that the better drone is able to do them (“Why?”). Visual explanations highlight risks from obstacles that users may have overlooked (“Where?”). A model induction user study showed that post-decision explanations produced a small effect on the participants’ abilities to identify the better of two imperfect drones and their plans for a mission, but they did not teach participants to judge the multiple success factors in complex missions as well as the AI pilots. In a decision support variation of the task, users would receive pre-decision explanations to help them to decide when to trust the XAI’s decision. In a fielded XAI application, every drone available for a mission may lack some competencies. We created a proof-of-concept demonstration of automatic ways to combine knowledge from multiple imperfect AIs to get better solutions that the individual AIs do not find on their own. This paper reports on the research challenges, technical approach, and findings of the project and also reflects on the multidisciplinary journey that we took.

This letter provides a retrospective analysis of our team’s research performed under the DARPA Explainable Artificial Intelligence (XAI) project. We began by exploring salience maps, English sentences, and lists of feature names for explaining the behavior of deep-learning-based discriminative systems, especially visual question answering systems. We demonstrated limited positive effects from statically presenting explanations along with system answers – for example when teaching people to identify bird species. Many XAI performers were getting better results when users interacted with explanations. This motivated us to evolve the notion of explanation as an interactive medium – usually, between humans and AI systems but sometimes within the software system. We realized that interacting via explanations could enable people to task and adapt ML agents. We added affordances for editing explanations and modified the ML system to act in accordance with the edits to produce an interpretable interface to the agent. Through this interface, editing an explanation can adapt a system’s performance to new, modified purposes. This deep tasking, wherein the agent knows its objective and the explanation for that objective will be critical to enable higher levels of autonomy.

As with almost all data, museum collection catalogues are largely unstructured, variable in consistency and overwhelmingly composed of thin records. The form of these catalogues means that the potential for new forms of research, access and scholarly enquiry that range across multiple collections and related datasets remains dormant. In the project Heritage Connector: Transforming text into data to extract meaning and make connections, we are applying a battery of digital techniques to connect similar, identical and related items within and across collections and other publications. In this paper we describe a framework to create a Linked Open Data knowledge graph (KG) from digital museum catalogues, connect entities within this graph to Wikidata, and create new connections in this graph from text. We focus on the use of machine learning to create these links at scale with a small amount of labelled data, on a mid-range laptop or a small cloud virtual machine. We publish open-source software providing tools to perform the tasks of KG creation, entity matching and named entity recognition under these constraints.

In the domain of Visual Question Answering (VQA), studies have shown improvement in users’ mental model of the VQA system when they are exposed to examples of how these systems answer certain Image-Question (IQ) pairs. In this work, we show that showing controlled counterfactual image-question examples are more effective at improving the mental model of users as compared to simply showing random examples. We compare a generative approach and a retrieval-based approach to show counterfactual examples. We use recent advances in generative adversarial networks (GANs) to generate counterfactual images by deleting and inpainting certain regions of interest in the image. We then expose users to changes in the VQA system’s answer on those altered images. To select the region of interest for inpainting, we experiment with using both human-annotated attention maps and a fully automatic method that uses the VQA system’s attention values. Finally, we test the user’s mental model by asking them to predict the model’s performance on a test counterfactual image. We note an overall improvement in users’ accuracy to predict answer change when shown counterfactual explanations. While realistic retrieved counterfactuals obviously are the most effective at improving the mental model, we show that a generative approach can also be equally effective.

Knowledge Graphs have been fast emerging as the {\em de facto} standard to model and explore knowledge in weakly structured data. Large corpora of documents constitute a source of weakly structured data of particular interest for both the academic as well as the industrial world. Key examples include scientific publications, technical reports, manuals, patents, regulations, etc. Such corpora embed many facts that are elementary to critical decision making or enabling new discoveries. In this paper, we present a scalable cloud platform to create and serve Knowledge Graphs, which we named \textit{Corpus Processing Service}. Its purpose is to process large document corpora, extract the content and embedded facts, and ultimately represent these in a consistent knowledge graph that can be intuitively queried. To accomplish this, we use state-of-the-art natural language understanding models to extract entities and relationships from documents converted with our previously presented CCS platform. This pipeline is complemented with a newly developed graph engine which ensures extremely performant graph queries and provides powerful graph analytics capabilities. Both components are tightly integrated and can be easily consumed through REST APIs. Additionally, we provide user-interfaces to control the data ingestion flow and formulate queries using a visual programming approach. The CPS platform is designed as a modular microservice system operating on Kubernetes clusters. Finally, we validate the quality of queries on our truly end-to-end knowledge pipeline in a real-world application in the oil and gas industry. To date, the capabilities of CPS are successfully leveraged in more than 5 client engagements.

Purpose: The field of cancer nanomedicine has made significant progress, but its clinical translation is impeded by many challenges, such as the difficulty in analysing intracellular anticancer drug release by the nanocarriers due to the lack of suitable tools. Here, we propose the development of a combinatorial imaging and analysis technique to evaluate anticancer drug such as doxorubicin HCl (DOX) released by a nanocarrier inside the HCT116 colon cancer cells and its subsequent intracellular accumulation. Procedure: Fluorescent cell images were captured and subjected to combined image analysis and machine learning based procedures to assess and quantify the delivery and retention rate of DOX inside the cancer cells by multifunctional CNT-DOX-Fe3O4nanocarrier. Results: We show that DOX in HCT116 cells was higher for multifunctional CNT-DOX-Fe3O4nanocarrierthan free DOX, indicating efficient and steady release of DOX as well as superior retentive property of the nanocarrier. Initially (1 h and 4 h) the luminance intensity of DOX in the cell cytoplasm delivered by CNT-DOX-Fe3O4nanocarrier was ~0.34 times and ~0.42 times lesser than that of free DOX delivered normally. However, at 24 h and 48 h post treatment the luminance intensity of DOX for CNT-DOX-Fe3O4nanocarrier was ~1.98 times and 1.92 times higher than that of free DOX. Furthermore, the luminance intensity of DOX for CNT-DOX-Fe3O4in the whole cell was ~1.35 times and ~1.62 times higher than that of free DOX at 24h and 48 h, respectively. Conclusions: The high-throughput nature of our image analysis workflow allowed us to automate the process of DOX retention analysis, and enabled us to devise machine learning-based modeling to predict the percentage of anticancer drug retention in cells. The development of models to automatically quantify and predict intra-cellular drug release in cancer cells could benefit personalized treatments by optimizing the design of nanocarriers.

We reflect on the progress in the area of Explainable AI (XAI) Program relative to previous work in the area of intelligent tutoring systems (ITS). A great deal was learned about explanation—and many challenges uncovered—in research that is directly relevant to XAI. We suggest opportunities for future XAI research deriving from ITS methods, as well as the challenges shared by both ITS and XAI in using AI to assist people in solving difficult problems effectively and efficiently.

Attention maps, a popular heatmap-based explanation method for Visual Question Answering (VQA), are supposed to help users understand the model by highlighting portions of the image/question used by the model to infer answers. However, we see that users are often misled by current attention map visualizations that point to relevant regions despite the model producing an incorrect answer. Hence, we propose Error Maps that clarify the error by highlighting image regions where the model is prone to err. Error maps can indicate when a correctly attended region may be processed incorrectly leading to an incorrect answer, and hence, improve users’ understanding of those cases. To evaluate our new explanations, we further introduce a metric that simulates users’ interpretation of explanations to evaluate their potential helpfulness to understand model correctness. We finally conduct user studies to see that our new explanations help users understand model correctness better than baselines by an expected 30% and that our proxy helpfulness metrics correlate strongly (rho>0.97) with how well users can predict model correctness.

The ability to predict the onset of labour is seen to be an important tool in a clinical setting. Magnetomyography has shown promise in the area of labour imminency prediction, but its clinical application remains limited due to high resource consumption associated with its broad number of channels. In this study, five electrode channels, which account for 3.3% of the total, are used alongside a novel signal decomposition algorithm and low complexity classifiers (logistic regression and linear-SVM) to classify between labour imminency due within 0–48hrs and >48hrs. The results suggest that the parsimonious representation comprising of five electrode channels and novel signal decomposition method alongside the candidate classifiers could allow for greater affordability and hence clinical viability of the magnetomyography-based prediction model, which carries a good degree of model interpretability.

The ease with which mobile money is used to facilitate cross-border payments presents a global threat to law enforcement in the fight against laundering and terrorist financing. This paper aims to use machine learning classifiers to predict transactions flagged as a fraud in mobile money transfers. Data for this paper came from real-time transactions that stimulate a  well-known mobile transfer fraud scheme. This paper uses logistic regression as the baseline model and compares it with ensembles and gradient descent models. The results indicate that the established logistic regression model did not perform too poorly compared to the other models. The random forest classifier had the most outstanding performance among all measures. The amount of money transferred was the top feature to predict money laundering transactions through mobile money transfers. These findings suggest that more research is needed to improve the logistic regression model. The random forest classifier should be further explored as a potential tool for law enforcement and financial institutions to detect money laundering activities in mobile money transfers.

The goal of continual learning (CL) is to learn a sequence of tasks without suffering from the phenomenon of catastrophic forgetting. Previous work has shown that leveraging memory in the form of a replay buffer can reduce performance degradation on prior tasks. We hypothesize that forgetting can be further reduced when the model is encouraged to remember the evidence for previously made decisions. As a first step towards exploring this hypothesis, we propose a simple novel training paradigm, called Remembering for the Right Reasons (RRR), that additionally stores visual model explanations for each example in the buffer and ensures the model has “the right reasons” for its predictions by encouraging its explanations to remain consistent with those used to make decisions at training time. Without this constraint, there is a drift in explanations and increase in forgetting as conventional continual learning algorithms learn new tasks. We demonstrate how RRR can be easily added to any memory or regularization-based approach and results in reduced forgetting, and more importantly, improved model explanations. We have evaluated our approach in the standard and few-shot settings and observed a consistent improvement across various CL approaches using different architectures and techniques to generate model explanations and demonstrated our approach showing a promising connection between explainability and continual learning. Our code is available at \url{https://github.com/SaynaEbrahimi/Remembering-for-the-Right-Reasons}

Benchmarking the performance of generative methods for drug design is complex and multifaceted. In this report, we propose a separation of concerns for de novo drug design, categorizing the task into three main categories: generation, discrimination, and exploration. We demonstrate that changes to any of these three concerns impacts benchmark performance for drug design taks. In this report we present Deriver, an open-source Python package that acts as a modular framework for molecule generation, with a focus on integrating multiple generative methods. Using Deriver, we demonstrate that changing parameters related to each of these three concerns impacts chemical space traversal significantly, and that the freedom to independently adjust each is critical to real-world applications having conflicting priorities. We find that combining multiple generative methods can improve optimization of molecular properties, and lower the chance of becoming trapped in local minima. Additionally, filtering molecules for drug-likeness (based on physicochemical properties and SMARTS pattern matching) before they are scored can hinder exploration, but can improve the quality of the final molecules. Finally, we demonstrate that any given task has an exploration algorithm best suited to it, though in practice linear probabilistic sampling generally results in the best outcomes, when compared to Monte Carlo sampling or greedy sampling. We intend that Deriver, which is being made freely available, will be helpful to others interested in collaboratively improving existing methods in de novo drug design centered around inheritance of molecular structure, modularity, extensibility, and separation of concerns.

Fast approximations of power flow results are beneficial in power system planning and live operation. In planning, millions of power flow calculations are necessary if multiple years, different control strategies, or contingency policies are to be considered. In live operation, grid operators must assess if grid states comply with contingency requirements in a short time. In this paper, we compare regression and classification methods to either predict multi-variable results, e.g., bus voltage magnitudes and line loadings, or binary classifications of time steps to identify critical loading situations. We test the methods on three realistic power systems based on time series in 15min and 5min resolution of one year. We compare different machine learning models, such as multilayer perceptrons (MLPs), decision trees, k-nearest neighbors, gradient boosting, and evaluate the required training time and prediction times as well as the prediction errors. We additionally determine the amount of training data needed for each method and show results, including the approximation of untrained curtailment of generation. Regarding the compared methods, we identified the MLPs as most suitable for the task. The MLP-based models can predict critical situations with an accuracy of 97-98% and a very low number of false negative predictions of 0.0-0.64%.

We present a focused analysis of user studies in explainable artificial intelligence (XAI) entailing qualitative investigation. We draw on social science corpora to suggest ways for improving the rigor of studies where XAI researchers use observations, interviews, focus groups, and/or questionnaires to capture qualitative data. We contextualize the presentation of the XAI papers included in our analysis according to the components of rigor described in the qualitative research literature: 1) underlying theories or frameworks, 2) methodological approaches, 3) data collection methods, and 4) data analysis processes. The results of our analysis support calls from others in the XAI community advocating for collaboration with experts from social disciplines to bolster rigor and effectiveness in user studies.

How would you describe the features that a deep learning model composes if you were restricted to measuring observable behaviours? Explainable artificial intelligence (XAI) methods rely on privileged access to model architecture and parameters that is not always feasible for most users, practitioners, and regulators. Inspired by cognitive psychology research on humans, we present a case for measuring response times (RTs) of a forward pass using only the system clock as a technique for XAI. Our method applies to the growing class of models that use input-adaptive dynamic inference and we also extend our approach to standard models that are converted to dynamic inference post hoc. The experimental logic is simple: If the researcher can contrive a stimulus set where variability among input features is tightly controlled, differences in response time for those inputs can be attributed to the way the model composes those features. First, we show that RT is sensitive to difficult, complex features by comparing RTs from ObjectNet and ImageNet. Next, we make specific a priori predictions about RT for abstract features present in the SCEGRAM dataset, where object recognition in humans depends on complex intra-scene object-object relationships. Finally, we show that RT profiles bear specificity for class identity, and therefore the features that define classes. These results cast light on the model’s feature space without opening the black box.

Predicting equipment failure is important because it could improve availability and cut down the operating budget. Previous literature has attempted to model failure rate with bathtub-formed function, Weibull distribution, Bayesian network, or AHP. But these models perform well with a sufficient amount of data and could not incorporate the two salient characteristics; unbalanced category and sharing structure. Hierarchical model has the advantage of partial pooling. The proposed model is based on Bayesian hierarchical B-spline. Time series of the failure rate of 98 Republic of Korea Naval ships have been modeled as hierarchical model, where each layer corresponds to ship engine, Engine type, and Engine archetype. As a result of the analysis, the suggested model predicted the failure rate of an entire lifetime accurately in multiple situational conditions, including the amount of prior knowledge of the engine.