17Q0203_SOW_w-Header.docx

N00173-17-Q-0203

STATEMENT OF WORK

The contractor will assist NRL’s team on a project pertaining to DARPA’s Explainable Artificial Intelligence (XAI) Program. In particular, NRL’s team will be responsible for serving as XAI’s Evaluation Team, which will involve conducting formal empirical studies on the Developer Team’s software deliverables. NRL’s team will also conduct novel research on XAI, a topic with which the contractor is intimately familiar.

The XAI Program was conceived as a reaction to recent, dramatic success of machine learning (ML) techniques, which has led to a torrent of AI applications. Continued advances promise to produce autonomous systems that will perceive, learn, decide, and act on their own. However, the effectiveness of these systems is limited by the machine’s current inability to explain their decisions and actions to human users. The DoD is facing challenges that demand more intelligent, autonomous, and symbiotic systems. XAI will be essential if future warfighters are to understand, appropriately trust, and effectively manage an emerging generation of artificially intelligent machine partners.

Given this, the XAI Program’s goals are to create a suite of ML techniques that:

· Produce more explainable models, while maintaining a high level of learning performance (prediction accuracy); and

· Enable human users to understand, appropriately trust, and effectively manage the emerging generation of AI partners.

New machine-learning systems will have the ability to explain their rationale, characterize their strengths and weaknesses, and convey an understanding of how they will behave in the future. The strategy for achieving that goal is to develop new or modified ML techniques that will produce more explainable models. These models will be combined with state-of-the-art human-computer interface techniques capable of translating models into understandable and useful explanation dialogues for the end user.

The contractor is particularly well informed on the subject of XAI. In particular, the contractor has designed, developed, and evaluated several XAI systems that use a variety of techniques of interest to generate explainable models and enable human users to interact with and query these models for explanations of their decision making.

The following is a description of the work to be performed: the contractor will attend meetings related to DARPA’s Explainable Artificial Intelligence (XAI) Program, provide advice to NRL’s XAI project team (which is responsible for managing the XAI Program’s Phase 2 evaluations), and conduct research on methods for explanation generation.

Tasks concerning providing guidance to NRL’s XAI Evaluation Team:

· Provide advice on approaches to XAI including input on what research objectives need to be addressed and what technical approaches are likely to succeed.

· Provide advice and reports on human learning and explanation to help inform approaches to XAI.

· Provide advice on methods for evaluating XAI.

· Provide recommendations of datasets and testbeds for XAI.

Tasks concerning original research on XAI:

· Explore the relationship between “explanations” generated by systems to justify decisions and explanations provided to systems to correct.

In particular, the contractor will conduct this research in the context of two types of problems. The first concerns the automated identification, from imagery, of biological specimens (e.g., birds) in which people can and do give explanations that distinguish similar entities. The second concerns medical diagnosis (e.g., skin cancer diagnosis), which is a task where explanations are of particular importance, particularly when treatments are risky or painful.

The primary deliverables will include:

1. A report on human learning from explanations to inform approaches on XAI.

2. A report on recommendations of datasets and testbeds for use in XAI evaluations.

3. A publication describing research on explanation-generation methods used in AI agents, for both “internal” explanations that are consumed by the agent and “external explanations that are provided to the human user. Evaluations will focus on the problems of automated identification of biological specimens and medical diagnosis.