A Constrained Optimization Approach to Preserving Prior Knowledge During Incremental Training

• Published in: IEEE transactions on neural networks Vol. 19; no. 6; pp. 996 - 1009
• Main Authors: Ferrari, Silvia, Jensenius, Mark
• Format: Journal Article
• Language: English
• Published: New York, NY IEEE 01.06.2008; Institute of Electrical and Electronics Engineers
• Subjects: Adaptation, Psychological; Adaptive control; Adaptive critics; Algorithms; Applied sciences; Artificial Intelligence; Artificial neural networks; Backpropagation; Biological neural networks; Computer science; control theory; systems; Constraint optimization; Constraints; control; Exact sciences and technology; exploration; Function approximation; Humans; incremental training; Interference; Knowledge; Knowledge acquisition; knowledge acquisition and retention; Learning - physiology; Mathematical analysis; Mathematical models; memory; Neural networks; Neural Networks (Computer); online learning; Optimization; Preserving; Programmable control; Scanning tunneling microscopy; sigmoidal neural networks; Training; Short term; Input output; Adaptive critics; memory; Flight; Updating; Function approximation; incremental training; Adaptive control; Adaptive method; Backpropagation algorithm; Backpropagation; Gradient; knowledge acquisition and retention; exploration; Knowledge acquisition; Minimization; control; Neural network; Long term; Constrained optimization; Supervised learning; sigmoidal neural networks; Equality constraint; interference; online learning

In this paper, a supervised neural network training technique based on constrained optimization is developed for preserving prior knowledge of an input-output mapping during repeated incremental training sessions. The prior knowledge, referred to as long-term memory (LTM), is expressed in the form o...