Research Manifesto

Vision

CCL pursues fundamental and applied research in the context of real-world problems and applications which (i) inform and constrain the research, and (ii) provide an opportunity to demonstrate new technology in working systems. CCL tackles problems that are "cognitive" in two respects:

the solution requires cognition: CCL systems are intelligent and embody a theory of learning and cognition (AI)
the solution aids cognition: CCL systems are designed with cognitive principles in order to interact naturally with humans (HCC)

Problem Statement

CCL research lies at the intersection of AI (artificial intelligence) and HCC (human-centered computing). CCL's focus is not cognitive science, it is cognitive computing, with an emphasis on computing and systems issues critical to tackling real-world, large-amounts-of-information, continuous-task, human-in-the-loop problems. It brings together relevant pieces of cognitive science and artificial intelligence with newer approaches from intelligent systems, machine learning, and human-centered computing. CCL projects use collaborative teams to tackle real-world problems that require basic (publishable) research coupled with applied (deliverable) R&D. CCL is interested in problems with the following characteristics:

Real World: There is a real-world need or "customer" for the technology, and real data or analog is available from the application (e.g., manufacturing plant, high school, intelligence analysis).
Human In The Loop: There is a need for intelligent cognitive assistance for humans performing complex tasks (e.g., decision support, education support).
Information Triage: There is a large amount of data. New data arrives in a continuous, incoming stream and must be triaged as it arrives (e.g., equipment sensors, news feeds, blogs).
Learning From Experience: A complete representation or model cannot be coded manually and must be learned and adapted "on the job" (e.g., troubleshooting operator dealing with an unexpected problem).

In everyday terms, the user may be "looking for a needle in a haystack" at the same time as s/he is "drinking from a firehose". For example:

an intelligence analyst may search for specific information in response to an RFI (request for information) in a database that is constantly being updated
a machine operator may watch for specific warning signs in a stream of data from hundreds of sensors
a network operator may flag potential intrusions by monitoring a stream of network data

Solution and Approach

CCL's approach to such problems has the following characteristics:

Goal-Driven Learning: The system identifies explicit goals for information required for a task, and using these goals to guide the learning process.
Asynchronous Reasoning and Learning: The system learns as it reasons, since there may not be time to "stop and think".
Metacognitive Control: The system reasons about itself, deciding what it needs to learn, which algorithms to use, when to respond and when to spend time learning.
Opportunism: The system uses deliberative processes to identify what it needs to learn, and is ready to make use of opportunities to carry out that learning.
Case-Based Reasoning: The system learns from experience using machine learning algorithms, adapting to unexpected situations and improving its performance over time.
Temporal and Textual Data: The system reasons with patterns of data over time. These data may be numerical or language-based.

Example Projects (w/ Problem Characteristics)

Project	Real world problem	Human in the loop	Information triage	Learning
Sacre Bleu: Content review and decision support for information assurance Funded by ARDA	Document releasability and routing	Intelligence analyst, disclosure officer (Decision support)	Large volume of documents and email	Learning by observing analysts at work
Game AI: Intelligent agents for interactive games Funded by DARPA	Real-time strategy games	Other players (Transfer learning)	Real-time monitoring of game environment and player actions	Case-based learning during real-time strategy and action selection
CaseBook: Educational technology to support problem-based learning for science education Funded by NSF	K-16 science education	Teacher, students (Learning support)	Pinpointing relevant information in texts and web	Learning students' knowledge needs during task
GeneTrek: Interpretation of gene microarray data by text mining MEDLINE	Microarray data analysis, Document classification	Biologist, epidemiologist (Analysis support)	1000s of genes and millions of abstracts	Functional classification