Connectionist Models of Neurons, Learning Processes, and Artificial Intelligence

Volume 2084 of the series Lecture Notes in Computer Science pp 215-222


Neocognitron-Type Network for Recognizing Rotated and Shifted Patterns with Reduction of Resources

  • Shunji SatohAffiliated withJapan Society for the Promotion of ScienceDepartment of Applied Physics, Graduate School of Engineering, Tohoku University
  • , Shogo MiyakeAffiliated withDepartment of Applied Physics, Graduate School of Engineering, Tohoku University
  • , Hirotomo AsoAffiliated withDepartment of Electrical and Communication Engineering, Graduate School of Engineering, Tohoku University

* Final gross prices may vary according to local VAT.

Get Access


A rotation-invariant neocognitron was proposed by authors for recognition of rotated patterns. In this paper, we propose a new network in order to reduce the number of cells for the same purpose. The new network is based on the rotation-invariant neocognitron in its structure and based on an idea of hypothesis and its verification in its process. In the proposed model the following two processes are executed: 1) making a hypothesis of an angular shift of an input supported by an associative recall network and 2) verification of the hypothesis realized by mental rotation of the input. Computer simulations show that 1) the new network needs less cells than the original rotation-invariant neocognitron and 2) the difference of recognition rates between the proposed network and the original network is very little.


rotation invariant neocognitron the number of cells associative recall mental rotation hypothesis and verification