A biologically motivated visual memory architecture for online learning of objects

• Published in: Neural networks Vol. 21; no. 1; pp. 65 - 77
• Main Authors: Kirstein, Stephan, Wersing, Heiko, Körner, Edgar
• Format: Journal Article
• Language: English
• Published: Oxford Elsevier Ltd 2008; Elsevier Science
• Subjects: Algorithms; Applied sciences; Artificial intelligence; Coding, codes; Computer science; control theory; systems; Discrimination Learning - physiology; Exact sciences and technology; Hierarchical feature extraction; Humans; Incremental and life-long learning; Information, signal and communications theory; Learning and adaptive systems; Learning vector quantization; Memory - physiology; Models, Neurological; Motivation; Neural Networks (Computer); Online Systems; Pattern recognition; Pattern Recognition, Visual - physiology; Photic Stimulation; Sampling, quantization; Signal and communications theory; Signal processing; Stability-plasticity dilemma; Telecommunications and information theory; Incremental and life-long learning; Learning vector quantization; Stability-plasticity dilemma; Hierarchical feature extraction; Short term; Data compression; Shape detection; Implementation; Learning; Inelasticity; Aspect oriented; Acquisition process; Plasticity; Learning algorithm; Target detection; Computer vision; Visual memory; Vector quantization; Memory architecture; Long-term memory; On-line systems; Pattern recognition; Object recognition; Object detection; Signal processing; Feature extraction; Hierarchical system; Edge detection; Artificial intelligence
• ISSN: 0893-6080; 1879-2782
• DOI: 10.1016/j.neunet.2007.10.005

We present a biologically motivated architecture for object recognition that is based on a hierarchical feature-detection model in combination with a memory architecture that implements short-term and long-term memory for objects. A particular focus is the functional realization of online and incremental learning for the task of appearance-based object recognition of many complex-shaped objects. We propose some modifications of learning vector quantization algorithms that are especially adapted to the task of incremental learning and capable of dealing with the stability-plasticity dilemma of such learning algorithms. Our technical implementation of the neural architecture is capable of online learning of 50 objects within less than three hours.