Deep learning has recently seen rapid development and received significant attention due to its state-of-the-art performance on previously-thought hard problems. However, because of the internal complexity and nonlinear structure of deep neural networks, the underlying decision making processes for why these models are achieving such performance are challenging and sometimes mystifying to interpret. As deep learning spreads across domains, it is of paramount importance that we equip users of deep learning with tools for understanding when a model works correctly, when it fails, and ultimately how to improve its performance. Standardized toolkits for building neural networks have helped democratize deep learning; visual analytics systems have now been developed to support model explanation, interpretation, debugging, and improvement. We present a survey of the role of visual analytics in deep learning research, which highlights its short yet impactful history and thoroughly summarizes the state-of-the-art using a human-centered interrogative framework, focusing on the Five W's and How (Why, Who, What, How, When, and Where). We conclude by highlighting research directions and open research problems. This survey helps researchers and practitioners in both visual analytics and deep learning to quickly learn key aspects of this young and rapidly growing body of research, whose impact spans a diverse range of domains.

Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers

We have recently seen many successful applications of recurrent neural networks (RNNs) on electronic medical records (EMRs), which contain histories of patients' diagnoses, medications, and other various events, in order to predict the current and future states of patients. Despite the strong performance of RNNs, it is often challenging for users to understand why the model makes a particular prediction. Such black-box nature of RNNs can impede its wide adoption in clinical practice. Furthermore, we have no established methods to interactively leverage users' domain expertise and prior knowledge as inputs for steering the model. Therefore, our design study aims to provide a visual analytics solution to increase interpretability and interactivity of RNNs via a joint effort of medical experts, artificial intelligence scientists, and visual analytics researchers. Following the iterative design process between the experts, we design, implement, and evaluate a visual analytics tool called RetainVis, which couples a newly improved, interpretable, and interactive RNN-based model called RetainEX and visualizations for users' exploration of EMR data in the context of prediction tasks. Our study shows the effective use of RetainVis for gaining insights into how individual medical codes contribute to making risk predictions, using EMRs of patients with heart failure and cataract symptoms. Our study also demonstrates how we made substantial changes to the state-of-the-art RNN model called RETAIN in order to make use of temporal information and increase interactivity. This study will provide a useful guideline for researchers that aim to design an interpretable and interactive visual analytics tool for RNNs.

RetainVis: Visual Analytics with Interpretable and Interactive Recurrent Neural Networks on Electronic Medical Records

High-quality immersive display technologies are becoming mainstream with the release of head-mounted displays (HMDs) such as the Oculus Rift. These devices potentially represent an affordable alternative to the more traditional, centralised CAVE-style immersive environments. One driver for the development of CAVE-style immersive environments has been collaborative sense-making. Despite this, there has been little research on the effectiveness of collaborative visualisation in CAVE-style facilities, especially with respect to abstract data visualisation tasks. Indeed, very few studies have focused on the use of these displays to explore and analyse abstract data such as networks and there have been no formal user studies investigating collaborative visualisation of abstract data in immersive environments. In this paper we present the results of the first such study. It explores the relative merits of HMD and CAVE-style immersive environments for collaborative analysis of network connectivity, a common and important task involving abstract data. We find significant differences between the two conditions in task completion time and the physical movements of the participants within the space: participants using the HMD were faster while the CAVE2 condition introduced an asymmetry in movement between collaborators. Otherwise, affordances for collaborative data analysis offered by the low-cost HMD condition were not found to be different for accuracy and communication with the CAVE2. These results are notable, given that the latest HMDs will soon be accessible (in terms of cost and potentially ubiquity) to a massive audience.

Immersive Collaborative Analysis of Network Connectivity: CAVE-style or Head-Mounted Display?

Visual analytics for machine learning has recently evolved as one of the most exciting areas in the field of visualization. To better identify which research topics are promising and to learn how to apply relevant techniques in visual analytics, we systematically review 259 papers published in the last ten years together with representative works before 2010. We build a taxonomy, which includes three first-level categories: techniques before model building, techniques during modeling building, and techniques after model building. Each category is further characterized by representative analysis tasks, and each task is exemplified by a set of recent influential works. We also discuss and highlight research challenges and promising potential future research opportunities useful for visual analytics researchers.

/pdf/a-survey-of-visual-analytics-techniques-for-machine-learning-up38fctryh.pdf

A survey of visual analytics techniques for machine learning

Machine learning (ML) models are nowadays used in complex applications in various domains such as medicine, bioinformatics, and other sciences. Due to their black box nature, however, it may someti ...

The State of the Art in Enhancing Trust in Machine Learning Models with the Use of Visualizations

Visual analytic tools combine the complementary strengths of humans and machines in human-in-the-loop systems. Humans provide invaluable domain expertise and sensemaking capabilities to this discourse with analytic models; however, little consideration has yet been given to the ways inherent human biases might shape the visual analytic process. In this paper, we establish a conceptual framework for considering bias assessment through human-in-the-loop systems and lay the theoretical foundations for bias measurement. We propose six preliminary metrics to systematically detect and quantify bias from user interactions and demonstrate how the metrics might be implemented in an existing visual analytic system, InterAxis. We discuss how our proposed metrics could be used by visual analytic systems to mitigate the negative effects of cognitive biases by making users aware of biased processes throughout their analyses.

Warning, Bias May Occur: A Proposed Approach to Detecting Cognitive Bias in Interactive Visual Analytics

People often rank and order data points as a vital part of making decisions. Multi-attribute ranking systems are a common tool used to make these data-driven decisions. Such systems often take the form of a table-based visualization in which users assign weights to the attributes representing the quantifiable importance of each attribute to a decision, which the system then uses to compute a ranking of the data. However, these systems assume that users are able to quantify their conceptual understanding of how important particular attributes are to a decision. This is not always easy or even possible for users to do. Rather, people often have a more holistic understanding of the data. They form opinions that data point A is better than data point B but do not necessarily know which attributes are important. To address these challenges, we present a visual analytic application to help people rank multi-variate data points. We developed a prototype system, Podium, that allows users to drag rows in the table to rank order data points based on their perception of the relative value of the data. Podium then infers a weighting model using Ranking SVM that satisfies the user's data preferences as closely as possible. Whereas past systems help users understand the relationships between data points based on changes to attribute weights, our approach helps users to understand the attributes that might inform their understanding of the data. We present two usage scenarios to describe some of the potential uses of our proposed technique: (1) understanding which attributes contribute to a user's subjective preferences for data, and (2) deconstructing attributes of importance for existing rankings. Our proposed approach makes powerful machine learning techniques more usable to those who may not have expertise in these areas.

Podium: Ranking Data Using Mixed-Initiative Visual Analytics

Visual analytics techniques help users explore high-dimensional data. However, it is often challenging for users to express their domain knowledge in order to steer the underlying data model, especially when they have little attribute-level knowledge. Furthermore, users' complex, high-level domain knowledge, compared to low-level attributes, posits even greater challenges. To overcome these challenges, we introduce a technique to interpret a user's drawings with an interactive, nonlinear axis mapping approach called AxiSketcher. This technique enables users to impose their domain knowledge on a visualization by allowing interaction with data entries rather than with data attributes. The proposed interaction is performed through directly sketching lines over the visualization. Using this technique, users can draw lines over selected data points, and the system forms the axes that represent a nonlinear, weighted combination of multidimensional attributes. In this paper, we describe our techniques in three areas: 1) the design space of sketching methods for eliciting users' nonlinear domain knowledge; 2) the underlying model that translates users' input, extracts patterns behind the selected data points, and results in nonlinear axes reflecting users' complex intent; and 3) the interactive visualization for viewing, assessing, and reconstructing the newly formed, nonlinear axes.

AxiSketcher: Interactive Nonlinear Axis Mapping of Visualizations through User Drawings

Recently, an approach for determining the value of a visualization was proposed, one moving beyond simple measurements of task accuracy and speed. The value equation contains components for the time savings a visualization provides, the insights and insightful questions it spurs, the overall essence of the data it conveys, and the confidence about the data and its domain it inspires. This articulation of value is purely descriptive, however, providing no actionable method of assessing a visualization's value. In this work, we create a heuristic-based evaluation methodology to accompany the value equation for assessing interactive visualizations. We refer to the methodology colloquially as ICE-T, based on an anagram of the four value components. Our approach breaks the four components down into guidelines, each of which is made up of a small set of low-level heuristics. Evaluators who have knowledge of visualization design principles then assess the visualization with respect to the heuristics. We conducted an initial trial of the methodology on three interactive visualizations of the same data set, each evaluated by 15 visualization experts. We found that the methodology showed promise, obtaining consistent ratings across the three visualizations and mirroring judgments of the utility of the visualizations by instructors of the course in which they were developed.

A Heuristic Approach to Value-Driven Evaluation of Visualizations

Visual analytic systems, especially mixed-initiative systems, can steer analytical models and adapt views by making inferences from users’ behavioral patterns with the system. Because such systems rely on incorporating implicit and explicit user feedback, they are particularly susceptible to the injection and propagation of human biases. To ultimately guard against the potentially negative effects of systems biased by human users, we must first qualify what we mean by the term bias. Thus, in this chapter we describe four different perspectives on human bias that are particularly relevant to visual analytics. We discuss the interplay of human and computer system biases, particularly their roles in mixed-initiative systems. Given that the term bias is used to describe several different concepts, our goal is to facilitate a common language in research and development efforts by encouraging researchers to mindfully choose the perspective(s) considered in their work.

Emily Wall

Papers

Warning, Bias May Occur: A Proposed Approach to Detecting Cognitive Bias in Interactive Visual Analytics

Podium: Ranking Data Using Mixed-Initiative Visual Analytics

AxiSketcher: Interactive Nonlinear Axis Mapping of Visualizations through User Drawings

A Heuristic Approach to Value-Driven Evaluation of Visualizations

Four Perspectives on Human Bias in Visual Analytics