Open Access. Powered by Scholars. Published by Universities.®
- Discipline
Articles 1 - 4 of 4
Full-Text Articles in Entire DC Network
Use Of Time Varying Dynamics In Neural Network To Solve Multi-Target Classification, S. N. Balakrishnan, J. Rainwater
Use Of Time Varying Dynamics In Neural Network To Solve Multi-Target Classification, S. N. Balakrishnan, J. Rainwater
Mechanical and Aerospace Engineering Faculty Research & Creative Works
Several types of solutions exist for multiple target tracking. These techniques are computation-intensive and in some cases very difficult to operate online. The authors report on a backpropagation neural network which has been successfully used to identify multiple moving targets using kinematic data (time, range, range-rate and azimuth angle) from sensors to train the network. Preliminary results from simulated scenarios show that neural networks are capable of learning target identification for three targets during the time period used during training and a time period shortly after. This effective classification period can be extended by the use of networks in coordination …
An Optical Adaptive Resonance Neural Network Utilizing Phase Conjugation, Donald C. Wunsch, D. J. Morris, T. P. Caudell, R. A. Falk
An Optical Adaptive Resonance Neural Network Utilizing Phase Conjugation, Donald C. Wunsch, D. J. Morris, T. P. Caudell, R. A. Falk
Electrical and Computer Engineering Faculty Research & Creative Works
An adaptive resonance (ART) device has been conceived that is fully optical in the input-output processing path. It is based on holographic information processing in a phase-conjugating crystal. This sets up an associative pattern retrieval in a resonating loop utilizing angle-multiplexed reference beams for pattern classification. A reset mechanism is used to reject any given beam, allowing an ART search strategy. The design is similar to an existing nonlearning optical associative memory, but it does allow learning and makes use of information the other device discards. This device is expected to offer higher information storage density than alternative ART implementations.
Semi-Supervised Adaptive Resonance Theory (Smart2), Christopher J. Merz, William E. Bond, Daniel C. St. Clair
Semi-Supervised Adaptive Resonance Theory (Smart2), Christopher J. Merz, William E. Bond, Daniel C. St. Clair
Computer Science Faculty Research & Creative Works
Adaptive resonance theory (ART) algorithms represent a class of neural network architectures which self-organize stable recognition categories in response to arbitrary sequences of input patterns. The authors discuss incorporation of supervision into one of these architectures, ART2. Results of numerical experiments indicate that this new semi-supervised version of ART2 (SMART2) outperformed ART for classification problems. The results and analysis of runs on several data sets by SMART2, ART2, and backpropagation are analyzed. The test accuracy of SMART2 was similar to that of backpropagation. However, SMART2 network structures are easier to interpret than the corresponding structures produced by backpropagation.
Moving Object Recognition And Guidance Of Robots Using Neural Networks, Abhijit Neogy, S. N. Balakrishnan, Cihan H. Dagli
Moving Object Recognition And Guidance Of Robots Using Neural Networks, Abhijit Neogy, S. N. Balakrishnan, Cihan H. Dagli
Mechanical and Aerospace Engineering Faculty Research & Creative Works
The design of a robust guidance system for a robot is discussed. The two major tasks for this guidance system are the online recognition of a moving object invariant to rotation and translation, and tracking the moving object using a neural-network-driven vision system. This system included computer software ported to the IBM PC and interfaced with an IBM 7535 robot. The operation of this guidance system involved recognition of a moving object and the ability to track it till the robot and effector was in close proximity of the object. It was found that the robot was able to track …