Deep Learning - Test & Evaluation (DL-T&E)

Project Description

Networks of neurons in human brains are multi-layered and deep. However, until recently, training feed-forward neural networks with more than three layers has been impractical. Development of deep learning (DL) approaches have made training of deep multi-layered networks feasible. Application of deep networks to tasks such as feature detection in images and natural language processing have produced promising results. However, their performance on more complex tasks remains ill-understood. DARPA is exploring the bounds of DL approaches by applying it to tasks such as feature detection in movies and performing sentiment analysis in text documents.

We are collaborating with the Navy Center for Applied Research in Artificial Intelligence (NCARAI) at the Naval Research Laboratory (NRL) to develop a platform for independently evaluating the novel techniques developed by the program participants. We are also standardizing evaluation metrics for cross comparison of new systems, and compiling the corpora for the evaluation.

Accomplishments

We've completed the development of the evaluation platform, which includes a library of evaluation metrics. We have also identified and annotated several multi-modal corpora for participant systems evaluation.