Autonomous Data Extractor Tool (ADE) Based On Real-Time and Adaptive Unsupervised Learning (RAUL)

T.A. Duong
Adaptive Computation LLC, United States

Keywords: Unsupervised learning, adaptation, autonomy, dynamic retraining, data training generator, streaming video data

Adaptive Computation LLC (ADC) proposes to develop and demonstrate a tool to rapidly and adaptively detect, recognize, crop, and extract images of objects of interest from streaming video, YouTube clips, or other sources. This autonomous data extractor (ADE) will automatically construct a database containing a large number of images of the object, in a wide variety of orientations, from raw video. The ADE tool will serve in two purposes: a) enabling users to satisfy the “hunger” of deep learning and reinforcement learning systems for massive sets of training data filling the “missing link” for intelligent system and b) providing on-line data in real-time for "Dynamic Retraining" that meet DoD mission needs. ADC proposes to base this tool on a Bio-Inspired Feature Extractor (BIFEE), based on the inspiration of a system that combines three patented elements: an Extended Visual pathway (EViP) model based on the mammalian visual system, and a Real Time Feature Extraction Algorithm (FEA) that extracts the bio-inspired shape features of objects. This is a novel and powerful bio-inspired technique that optimally detects moving or stationary objects in a heterogeneous dynamic environment.