6

Session 5: Learning from Noisy, Adversarial Inputs

OPENING REMARKS

David Honey, Office of the Director of National Intelligence

David Honey, Office of the Director of National Intelligence (ODNI), thanked everyone for participating in this experiment to change the way that ODNI does business and operations for research and development planning in its science and technology office. He stated that ODNI’s main function is intelligence integration; in other words, how do the people and processes of the various agencies work together more effectively to meet national security needs?

Honey described the Intelligence Community Information Technology Enterprise (IC ITE)¹ initiative, which will move the community to a cloud-based environment where the various information systems, databases systems, and tools will coexist and promote more effective sharing. Honey initially considered this a bold step because ODNI decided to move forward with this initiative before the relevant technology had been developed. Because this approach is common in government acquisitions, he noted that smart decisions about research and development investments are crucial.

Although the intelligence community has a wide range of problems and varied members, one commonality, Honey stated, is that all of its activities revolve around collection and analysis. And collection can be complicated because the more that is collected, the larger the big data problem becomes. Thus, problems have to be considered holistically. Honey suggested that ODNI look at how the technology, the people, the processes, and the system all work together and improve the overall architecture instead of simply attempting to improve individual components. Generating partnerships and building frameworks helps ODNI to determine what research and development investments need to be made in research related to national security.

___________________

¹ For more information about IC ITE, see Office of the Director of National Intelligence, “IC IT Enterprise,” fact sheet, https://www.dni.gov/files/documents/IC%20ITE%20Fact%20Sheet.pdf, accessed August 10, 2017.

HARNESSING MACHINE LEARNING FOR GLOBAL DISCOVERY AT SCALE

Mikel Rodriguez, MITRE

Mikel Rodriguez, MITRE, focused his presentation on harnessing machine learning (especially computer vision) to do global discovery at scale. He explained that because there is both an explosion of data and the potential for conflicts to arise without warning, it is imperative to develop new ways to collect intelligence. Rodriguez discussed the revolution in commercial imaging: today there are more pixels, different modalities, and more imaging companies. There is also an abundance of uncooperative and nontraditional sources (e.g., social media data) that could be harnessed to better focus our attention and our government assets, according to Rodriguez.

Although visual data may be considered the biggest form of big data, it was previously challenging to access it, index it, effectively reason about it, or move from pixels to meaning when comparing two images. These problems of representation began to be addressed with deep learning approaches, thanks to the work of Yann LeCun, and with the presence of graphical processing units.

Rodriguez commented that this thriving research field that started in academia quickly spread to industry, national laboratories, and federally funded research and development centers. Deep learning and computer vision algorithms can be used to detect images, understand patterns, and make insights about objects or events of interest. Particular attention is focused on analyzing social media data, though this is an especially noisy source, Rodriguez explained.

With the success of deep learning systems, people wonder about the role of computer vision in the future. As a community, Rodriguez suggested further investment in the areas of sensitive analytics and semantic segmentation. He noted the issue of “long tails” (i.e., the many specific or rare objects that are difficult to model because so few examples exist) in the visual world and the possibility of mixing synthetic data with real examples to create better classifiers. Rodriguez explained that labeling data is expensive: it is irrational to expect that large groups of people will be available to label all of the data the intelligence community has. Instead, it would be more effective to try to use current data sets to train algorithms.

Rodriguez said that deployment of deep learning approaches remains a challenge in certain domains. For example, deep learning approaches currently depend on expensive graphical processing units that are computationally powerful but draw much energy. While less expensive hardware could be used for object detection, Rodriguez noted that there is a balance between how much recognition power one needs and how much electrical power one is drawing from a device.

In conclusion, Rodriguez summarized the key points of his presentation: (1) Deep learning has recently led to tremendous progress in representation learning when large amounts of labeled data are available; (2) Image representation is powerful and has led to breakthroughs in classification, segmentation, and scene understanding; (3) Deep learning does not necessarily equate to deep insight; and (4) More work needs to be done on complex activities, temporal sequences, rare targets, and modalities unique to the Department of Defense/intelligence community, as well as counter artificial intelligence assurance.

Devanand Shenoy, Department of Energy, asked how often deep learning architectures are used in image processing. Rodriguez emphasized that there is much interesting work in other types of architectures, though convolutional neural networks currently work the best. He noted that neuromorphic computing is exciting but has yet to materialize. Rodriguez suggested returning to the basics to determine what should be optimized and how it should be built. Anthony Hoogs, Kitware, Inc., noted that the workshop had yet to cover the influences of deep learning and convolutional neural networks on image processing, though they have been substantial. For the intelligence community in particular, there are improved approaches for image clean-up, for example.

Josyula Rao, IBM, asked which approaches should be used when the number of images is very small and specialized and how deep learning evolves to address general problems in security. He also asked Rodriguez to provide more details on what is happening in the space of counter artificial intelligence assurance since adversaries are using artificial intelligence tools for synthesizing malware that will evade security controls. Rodriguez explained that there is a lot of work being done both on the offensive and defensive sides. He said that it is important for the intelligence community to recognize what is possible—for example, it is easy for adversaries to extract information

from a model on which one has trained data. Rodriguez also remarked that there is not currently much economic incentive to work in this area, so additional investment would help to address the remaining gaps.

Pelayo Fernandez, Department of Defense, asked if synthetic data from gamers could be useful in helping algorithms learn better varieties of targets (as opposed to relying on labelers). Rodriguez commented that although there is a long history of using synthetic data in the computer vision community, he cautioned against relying on it too much, because it remains difficult to simulate the real world. Hoogs suggested that the combination of synthetic data and adversarial networks could be powerful. He also noted that the gaming community has three-dimensional models of Russian military equipment that are openly available.

Ellen Voorhees, National Institute of Standards and Technology, added that, in her experience, synthetic data does not capture the real world or the problem at hand adequately. Kathy McKeown, Columbia University, suggested that as we move to using multi-modal data, there may be cases where there are images with accompanying text that could be useful for generating captions and descriptions. Rodriguez agreed that this is an appropriate area for future research but noted the challenges that will arise between siloes that do not communicate with one another.

Rama Chellappa, University of Maryland, College Park, noted that it is difficult, if not impossible, to gain access to restricted data, though data sets for action recognition are becoming more accessible. Some sensitive data will never be available, so that challenge continues. He also explained that although synthetic data has its faults, it is still worth using as long as the limitations are understood. In reference to McKeown’s earlier comment on multi-modal data, Hoogs added that there is a treasure trove of data that the intelligence community could leverage—for example, people blogging in real time while watching unmanned aerial vehicle camera footage.