I developed this simulator of the Neocognitron neural network as a part of my master thesis. Neocognitron was initially suggested by his author, prof. Kunihiko Fukushima, as a hierarchical multi-layered model of mammalian vision. Neocognitron has been successfully tested in character recognition tasks.
The simulator offers a flexible software tool for probing Neocognitron in action. User can customize a structure of the network and control its training with arbitrary image data. Then, the designed network can be tested in pattern recognition task. The simulator can be also very helpful for understanding basic processes occurring in the biological visual systems.
The present simulator, called Beholder 2, is based on the advanced version of Neocognitron model, described in "Neocognitron capable of incremental learning", K.Fukushima (2004). New network features, such as incremental learning, synaptic plasticity and structural self-organization, greatly improve the network recognition performance. The simulator is explicitly driven by the theoretical model.
I added several features that simplify deeper exploration and understanding of model work, e.g. input segmentation and visualization of the network’s prototype features.
This software simulator is FREEWARE. It may be uses for any education and research purposes. It may be freely duplicated and distributed, however, the author holds a copyright on. It must not be distributed or used for any commercial purpose.
The simulator was designed with Borland Delphi 6.0 (c)Borland Software Corporation.