Paper Overview
This paper presents an unsupervised neural network models for visual pattern recognition. The network is not recognize the influence of the position of the object, to complete the identification by the method of geometric similarity. The thesis of such a network is called neocognitron.
Misalignment and most seriously affected by the neural network input mode shape distortion, that is, the same pattern presented in a different location or a different level of the same sleek style, the traditional neural networks as a different mode. However, self-organizing neural network model proposed in response to the network mode is hardly affected by the location of the stimulus.
Neural Network Architecture
Hubel and Wiesel cells had been a Category: LGB (lateral geniculate body) → simple cells → complex cells → lower order hypercomplex cells → higher order hypercomplex cells
As shown in FIG. 1, neocognitron cascade-connected by a series of modular configuration, until all the structure of an input layer is U0. Each module structures are connected by two layers of cells cascade. A first module layer "S cells (S-cells)" Composition, S-cells or simple cells corresponding to the lower order hypercomplex cells, we call layer s and layer s l module is expressed as Usl . The second layer is composed of module "C cells (c-cells)", corresponding to complex cells or higher order hypercomplex cells. We call it the layer c, c and l-th layer is represented as module Ucl. In this neural network, only the input cell layer synaptic plasticity s and modifiability.
S one of the cells or cell c based on different parts of the best feel its upper input stimulus, it is divided into subgroups (subgroups). Since each subgroup of cells can be delineated as a two-dimensional graphics, so we call this subgroup as "flat-cell (cell-plane)". S-plane and C-plane are represented by a cell plane and cell c s cells.
FIG 2 is a schematic diagram interlayer interconnected. Each rectangular with a bold line drawn representing the s-plane or a c-plane, each of the thin line drawn vertically represent a quadrangular s layer or layers c, wherein c s layer or layers are closed.
The total number of cells in each cell plane in the plane of the network increases as the depth of the cells is reduced. In the last module, each of the C-cell becomes very large accepted domain that covers the entire area of the input layer, and each of the C-plane is determined to be in a C-cell.
Self-organization network
First, the stimulation pattern at each occurrence, selected several "representative (Representative) 's of the S cells from each layer. Representative S cells are selected out of a large amount of output S cells, each representing a selected plane at most. Similar procedure selected from S cells in the conventional recognition unit (conventional cognitron) enhancement process selected cells.
Representative synaptic input S cells is enhanced with the rms-type in the same manner. On the plane S, if the cell is selected, the other cell synapse input S on a plane will be strengthened in the same manner. If no cells are selected on the S plane, the synaptic input of all cells in the S plane will not be strengthened.
Works Network
In this network, the input pattern and standard patterns previously obtained from the study are compared. This comparison is performed not by a large window in pattern matching is performed directly, but rather by the small segment pattern matching window. Only when the difference between the two modes in any small window does not exceed a certain limit, the network will judge these patterns are consistent with other modes.
Comparing each stage, the mode change of the position tolerance is gradually increased. Compare the size of the window in a higher stage will become greater. In the last phase, the window is large enough, the entire input mode information can be observed simultaneously.
Since the pattern matching the first stage is tested in parallel in a plurality of small windows, so long as the first phase through the screening of small errors it can be considered that they are matched. Thus, even if the input pattern has some distortion in the shape of a network it is possible to make a correct pattern recognition.
