Selective Visual Attention: Computational Models and Applications

5.4 Attention with Memory of Learning and Amnesic Function

As described above, visual memory plays a major role in top-down knowledge storage and the control of visual selective attention. In the model mentioned above, the top-down information is stored in the visual memory, working memory (WM) and long term memory (LTM), in the form of a decision tree. The working memory is similar to short-term memory (STM) that helps to match the target located at the current attention focus in the image by search of the visual memory with the decision tree. However, most models with working memory mainly compute selective attention in a static image [14, 24, 25, 50]. In some applications, such as walking robot vision which needs to accommodate huge amounts of video data over a long period, the visual memory with online top-down knowledge learning and fast retrieving becomes more important.

A perfect decision tree of classification and regression for high dimension data referred to as hierarchical discriminant regression (HDR) is proposed in [51]. The HDR tree does not require any global distribution assumption, and has high recognition precision and fast execution speed, which has been used in developmental robots [52]. Later, the HDR tree with online learning was developed by the same authors, and called an incremental hierarchical discriminant regression (IHDR) tree in [53]. The HDR and IHDR trees are very successful in pattern recognition and data classification, but it is not considered ...