Rotation Invariant Neocognitron

• Main Author: TING, Christopher Hian Ann
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 1991
• Subjects: Physical Sciences and Mathematics
• Online Access: https://ink.library.smu.edu.sg/lkcsb_research/780; https://doi.org/10.1109/IJCNN.1991.170717
• Institution: Singapore Management University

A biologically inspired method that would reduce the computation load of the neocognitron supervised model is proposed. This method also provides a way to extend the neocognitron's invariance properties to include rotation invariance. The key idea is to incorporate a magnocellular pathway into Fukushima'a algorithm. A basic shift in the paradigm is that an input is said to be recognized when and only when one of the winners of the magnocellular pathway is validated by the parvocellular pathway. This is a modification of the original model where only the activities of the (parvocellular) grandmother cells indicate recognition. The rotation invariance comes easily by making the magnocellular pathway produce an orientation winner for each grandmother cell, exploiting the Zk-symmetry in the cell-planes of the first stage. Then, with another round of competition, the magnocellular pathway submits a few hypotheses for validification. The author has implemented this method on transputers. The simulation program can recognize numerals in arbitrary orientation