You are here

Robust, Adaptive Machine Learning (RAM)

Description:

TECHNOLOGY AREA(S): Space Platforms 

OBJECTIVE: Develop robust, adaptive machine learning capabilities that maintain and improve performance even as the data and behavior being modeled evolve to support decision making processes. 

DESCRIPTION: Enabling rapid Situational Awareness (SA) in support of (Battle Management Command and Control (BMC2) is critical for analyst assessment and commander decision-making from multi-INT data in complex and evolving multi-domain situations. Machine learning technology has been applied to support intelligence needs by automatically processing massive and increasing amounts of intelligence data to provide analysts with awareness and recognition of evolving activities, alerts, and actionable intelligence. However, existing machine learning-based systems that are largely static and non-interactive are not capable of adapting to changes in their input data streams or in the real-world behavior they are supposed to model. Adaptive systems are critical for enhancing decision-making posture. Recent research in machine learning provides us with hope that robust, adaptive machine learning systems can be developed. These include: statistical models for entity, relationship, and descriptor extraction; unsupervised techniques for topic and group discovery from text; automated structure learning; latent variable models that enable inferring joint structure for image/video and text; and active learning with user feedback. Evolving these capabilities to the next level requires developing an architecture that supports full-scale and robust adaptive machine learning across multi-INT/multi-domain data. Fully implementing such a capability requires addressing several challenges: managing model drift and uncertainty, maintaining and managing multiple complementary and possibly competing models, and automatically monitoring model performance to determine when a model needs to be updated, replaced, or retired, or when to solicit human feedback. This topic is seeking new machine learning architectures and technologies to support development of robust and adaptive machine learning based systems that can continue to operate when faced with noisy, diverse data-sets, changing data streams and/or evolving behavior, adapt through continuous learning, respond to dynamic environments and changes in adversary tactics, new mission requirements, and unforeseen contingencies, and development of machine learning algorithms capable of incremental model updating based on user feedback and/or data drift indicators. This topic seeks to bring state of the art machine learning and reasoning algorithms to: - Enable rapid situational awareness in support of Space BMC2 using multi-INT/multi-domain data - Increase Space Operator/Analyst and Senior leadership confidence in situational understanding of the space domain and in identification, evaluation, and selection of appropriate courses of action - Produces situational awareness indications and warnings of anomalous and threats associated with objects’ and entities’ activity behaviors - Support decision making processes under uncertain, changing, and time-sensitive conditions - Increase threat warning time, decrease forensic analysis time, increase accuracy of threat identification and characterization and support predictive analysis It is not anticipated that the government will provide GFE/GFI (including data) during Phase 1. It is anticipated that the government will provide access to data during Phase 2. 

PHASE I: Design a robust, adaptive machine learning architecture. Define a set of metrics for assessing performance. Deliverables will include a system architecture design, block diagram identifying data flows and interfaces, and identification of required data sets and demonstration use cases. 

PHASE II: Develop and demonstrate a prototype system based on the architecture defined in Phase I. Develop and implement a plan to test and measure the performance of the system against real data. The Phase II system should be tested on at least two use cases identified during Phase 1. 

PHASE III: RAM technologies/architectures will support a broad range of intelligence and intelligence-related applications. Commercial -The improved learning capability will allow rapid deployment and robust performance for a variety of business intelligence and marketing applications. 

REFERENCES: 

1: Sudderth, E., A. Torralba, W. Freeman, A. Willsky (2005). Describing visual scenes using transformed Dirichlet processes. NIPS, Vancouver, BC.

2:  Wang X., N. Mohanty, A. McCallum (2006). Group and topic discovery from relations in text. NIPS 2006.

3:  Jain, Vidit, E. Learned-Miller, A. McCallum (2007). People-LDA: anchoring topics to people using face recognition. International Conference on Computer Vision.

4:  Kemp C., J. Tenenbaum (2008). The discovery of structural form. PNAS 105:31

KEYWORDS: Machine Learning, Adaptive Learning, Robust Learning, Statistical Modeling 

CONTACT(S): 

Carolyn Sheaff (SMC/SYE) 

(315) 330-7147 

carolyn.sheaff@us.af.mil 

US Flag An Official Website of the United States Government