One could be excused for assuming that deep learning had or will soon usurp all credible work in reasoning, artificial intelligence and statistics, but like most ‘meme’ class broad generalizations the concept does not hold up to scrutiny. Memes don’t generally matter since the experts will always know better but in the case of Bayesian software like Stan and PyMC3 even its developers and advocates bemoan the apparent dominance of deep learning as manifested in popular culture, breathtaking performance and most problematically from funding agency peer review that impacts our ability to further advance the field. The facts however do not support the assumed dominance of deep learning in science upon closer examination. This letter simply makes the argument by the crudest of possible metrics, citation count, that once Computer Science is subtracted, Bayesian software accounts for nearly a third of research citations. Stan and PyMC3 dominate some fields, PyTorch, Keras and TensorFlow dominate others with lots of variation in between. Bayesian and deep learning approaches are related but very different technologies in goals, implementation and applicability with little actual overlap so this is not a surprise. While deep learning is backed by industry behemoths (Google, Facebook) the Bayesian efforts are not and it would behoove funders to recognize the impact of Bayesian software given its centrality to science.

Automatic video understanding is becoming more important for applications where real-time performance is crucial and compute is limited. Yet, accurate solutions so far have been computationally intensive. We propose efficient models for videos - Tiny Video Networks - which are video architectures, automatically designed to comply with fast runtimes and, at the same time are effective at video recognition tasks. The Tiny Video Networks run at faster-than-real-time speeds and demonstrate strong performance across several video benchmarks. These models not only provide new tools for real-time video applications, but also enable fast research and development in video understanding. Code and models are available.

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 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}

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.

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.

We study a user-guided approach for producing global explanations of deep networks for image recognition. The global explanations are produced with respect to a test data set and give the overall frequency of different “recognition reasons” across the data. Each reason corresponds to a small number of the most significant human-recognizable visual concepts used by the network. The key challenge is that the visual concepts cannot be predetermined and those concepts will often not correspond to existing vocabulary or have labelled data sets. We address this issue via an interactive-naming interface, which allows users to freely cluster significant image regions in the data into visually similar concepts. Our main contribution is a user study on two visual recognition tasks. The results show that the participants were able to produce a small number of visual concepts sufficient for explanation and that there was significant agreement among the concepts, and hence global explanations, produced by different participants.

Neural network architectures are achieving superhuman performance on an expanding range of tasks. To effectively and safely deploy these systems, their decision-making must to be understandable to a wide range of stakeholders. Methods to explain 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 any XAI method into four components: (1) the inference to be explained, (2) the explanatory medium, (3) the explainee model, and (4) the explainer model. The abstraction afforded by Bayesian Teaching to decompose any XAI method 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.

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.

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%.