Autonomous Intelligent Architecture—Preliminary Results

T.A. Duong and Q.D. Duong
Adaptive Computation LLC,
United States

Keywords: autonomous intelligence, short-term memory, long-term memory, self-extraction training data, self-learning, on-line feature adaptation, extended visual pathway


Abstract—We present an innovative architecture for the autonomous intelligent learning system, which comprises the feedback loop between unsupervised learning based on Extended Visual Pathway (EViP) firstly introducing in the book chapter [ ] to self-extract the on-line training data and dynamic supervised learning [ ] to digest active data by self-learning capability, to enable the autonomous intelligent learning system for dynamic perception and cognition assessment. Unsupervised learning EViP is viewed as short-term memory, while dynamic supervised learning is serving as long-term memory and the interactive feedback loop between them to facilitate the dynamic knowledge in the open environment. The primary results that are based on the video stream in the real-world field are reported. For short-term memory it requires detecting, recognizing, tracking and adapting the moving object and processed 5 frames/sec (self-extracting training data) while it took 5.8 mins to learn 7776 object samples (67x71 pixel array) (self-learning). With the equipped knowledge neural network processor, we tested with 1/3 of training and 2/3 of un-training data with full frame resolution (576x696) from raw video. The object recognition is 88.5% correct with the ranks 1.