Explaining decisions is an integral part of human communication, understanding, and learning, and humans naturally provide both deictic (pointing) and textual modalities in a typical explanation. The challenge is to build deep learning models that are also able to explain their decisions with similar fluency in both visual and textual modalities (see Fig. 2). Previous machine learning methods for explanation were able to provide a text-only explanation conditioned on an image in context of a task, or were able to visualize active intermediate units in a deep network performing a task, but were unable to provide explanatory text grounded in an image.
Existing approaches for deep visual recognition are generally opaque and do not output any justification text; contemporary vision-language models can describe image content but fail to take into account class-discriminative image aspects which justify visual predictions.
Hendriks et al.  propose a new model (see Fig. 3) that focuses on the discriminating properties of the visible object, jointly predicts a class label, and explains why the predicted label is appropriate for the image. The idea relies on a loss function based on sampling and reinforcement learning, which learns to generate sentences that realize a global sentence property, such as class specificity. This produces a fine-grained bird species classification dataset, and shows that an ability to generate explanations which are not only consistent with an image but also more discriminative than descriptions produced by existing captioning methods.
Although, deep models that are both effective and explainable are desirable in many settings, prior explainable models have been unimodal, offering either image-based visualization of attention weights or text-based generation of post-hoc justifications. Park et al.  propose a multimodal approach to explanation, and argue that the two modalities provide complementary explanatory strengths.
Two new datasets are created to define and evaluate this task, and use a model which can provide joint textual rationale generation and attention visualization (see Fig. 4). These datasets define visual and textual justifications of a classification decision for activity recognition tasks (ACT-X) and for visual question answering tasks (VQA-X). They quantitatively show that training with the textual explanations not only yields better textual justification models, but also better localizes the evidence that supports the decision.
Qualitative cases also show both where visual explanation is more insightful than textual explanation, and vice versa, supporting the hypothesis that multimodal explanation models offer significant benefits over unimodal approaches. This model identifies visual evidence important for understanding each human activity. For example to classify “mowing lawn” in the top row of Fig. 4 the model focuses both on the person, who is on the grass, as well as the lawn mower. This model can also differentiate between similar activities based on the context, e.g.“mountain biking” or “road biking.”
Similarly, when asked “Is this a zoo?” the explanation model is able to discuss what the concept of “zoo” represents, i.e., “animals in an enclosure.” When determining whether the water is calm, which requires attention to specific image regions, the textual justification discusses foam on the waves.
Visually, this attention model is able to point to important visual evidence. For example in the top row of Fig. 2, for the question “Is this a zoo?” the visual explanation focuses on the field in one case, and on the fence in another.
There are also other approaches to explanation that formulate heuristics for creating what have been called “Deep Visual Explanation” . For example, in the application to debugging image classification learned models, we can create a heat map filter to explain where in an image a classification decision was made. There are an arbitrary number of methods to identify differences in learned variable distributions to create such maps; one such is to compute a Kullback-Leibler (KL) divergence gradient, experiments with which are described in , and illustrated in (see Fig. 5). In that figure, the divergence for each input image and the standard VGG image classification predictor is rendered as a heat map, to provide a visual explanation of which portion of an image was used in the classification.