This paper describes initial progress in deep learning capable not only of fine-grained categorization tasks, such as whether an image of bird is a Western Grebe or a Clark's Grebe, but also explaining contrasts to make them understandable. Knowledge-discovery in databases has been described as the process of identifying valid, novel, potentially useful, and ultimately understandable patterns in data [1]. In spite of this, much of machine learning has focused on “valid” and “useful” with little attention paid to “understandable” [2- 6]. Recent work in deep learning has showed remarkable accuracy on a wide range of tasks [7], but produces models that are more difficult to interpret than most earlier approaches to artificial intelligence and machine learning. Our ultimate goal is to learn to annotate images to explain the difference between contrasting categories as found in bird guides or medical books.
HISTORY
The first author’s research on learning explainable models
from data started in the mid-1990s after interacting with
doctors on models for medical diagnosis [
BIRD IDENTIFICATION: A PILOT STUDY.
We prepared images of 12 birds divided into 6 sets of contrasting birds (e.g., Spotted Towhee and Eastern Towhee). Images were shown to four experienced bird watchers who were asked a yes-or-no question about the bird identification (e.g., “Is this a Spotted Towhee?”). Using an eye tracking system, we recorded the parts of the image that received attention. Figure 3 shows an example of where one subject focused on the wing of eastern towhee and spotted towhee, two similar pictures distinguished in part due to the spots on the wing. In contrast to distinguish a Clark’s Grebe from a Western Grebe that subject concentrated on the area around the eye and the bill.
Although suggestive of how bird watchers learn and attend to field marks to distinguish similar species, the data is preliminary and requires more subjects and statistical tests. Experiments are planned to show two contrasting images simultaneously to experienced bird watchers and to track attentional changes in novices as they learn to identify birds.
LEARNING DISCRIMINATIVE REGIONS OF
CONTASTING CATEGORIES
Deep learning for image classification has shown great
results [
We demonstrate the approach with a fine-grained
classification task of birds. We seek to find out which regions
of a bird image are the most important to distinguish it from
images of the most similar contrasting class of birds. To do
so, we train a known deep network [
To find these image regions, we backward propagate an output vector consisting of +1 for the most likely class and 1 for the next most likely class. This propagation is similar to the “backprop” neural network algorithm used to train deep networks, but in this case applied for explanation and not for training network weights.
This process identifies the most important pixels, i.e., those for which changes will cause the network to assign the given image to the most-similar (i.e., the next most likely) class instead of the correct class. This raw set of pixels is too sporadically distributed to provide a human-consumable explanation. To get larger coherent regions, we convolve windows with different sizes on the image and record the maximum change in each window. The windows with the maximum changes are the regions that are mostly contributing to misguide the network and the most important regions to explain the features that discriminate the contrasting categories. The process is depicted in Figure 5. As one example, the network correctly identified a test image as a cerulean warbler and the second most likely classification as a black throated blue warbler Figure 6 (left) highlights the regions of the image that were found to be most important in distinguishing the two classes including the eye and throat. Figure 6 (right) is an image of a black throated blue warbler which shows the difference in throat and neck.
To illustrate the impact of having a contrasting category, Figure 7 shows the regions that contribute most to producing the correct category without regard to finding the difference between a contrasting categories. (i.e. not propagating -1 for the next most similar class). Note that neither the throat nor eye are highlighted.
CONCLUSION We have begun to explore how machine learning may emulate how humans explain contrasting categories. Preliminary data show the features that experienced bird watchers use to differentiate contrasting categories. A network architecture learned similar features. In future work, we will explore how to label differentiating features using techniques similar to [9] with an ultimate goal to automate explanations similar to those found in bird guides and medical books.
ACKNOWLEDGEMENTS This was developed with funding from the DARPA Explainable AI Program under a contract from NRL. The views, opinions, and/or findings expressed are those of the author and should not be interpreted as representing the official views or policies of the DoD or the U.S. Government. 3.