Associative-Projective Neural Networks (APNNs) - Neural Networks and Micromechanics - page 76

Fig. 5.3 The associative field
structure
field. One of the set inputs of each neuron leaves the limits of the field and serves for
information input into the associative field (inputs S ).
We emphasize that the neurons of the associative field are connected with each
other with the aid of associative connections, whose weights can change in the
process of training. This is important because two additional types of connections
will be introduced below: projective and receptive.
Similarly to the set inputs, one of the inhibitory (reset) inputs of each neuron
leaves the field and serves for the forced inhibition of excited neurons (inputs R ).
The field has one common input to inhibit all the neurons of the field and the
synchronizing input C . All the additional outputs of neurons ( Q * ) inside the field are
supplied to the inputs of the field activity regulator ( FAR ), which determines a
quantity of excited neurons ( bn ) and establishes the threshold th (common for all
neurons) in such a way that the number of active neurons would not exceed a certain
value m , which is input into the associative field on the input M.
Let us consider the work of the associative field in detail by first examining its
timing diagram (Fig. 5.4 ). The threshold for the associative field neurons must be
selected in such a way that no more than m neurons in the network will be active. It
is achieved in one cycle of the work of the associative field. The duration of the
syncronizing signal is selected such that during its first part (high level of the sync
signal), the process of summing up the input signals will be completed. The field
activity regulator (FAR) begins to work along the back edge of the sync signal. In
this case, the outputs Q of all neurons of the field are locked and do not change
because they can change only along the front edge of the sync signal. The FAR
smoothly decreases the threshold th (Fig. 5.4 ) for all neurons of the field and
correspondingly changes signals at the additional outputs Q * so that a quantity of
active neurons (marked bn in Fig. 5.4 ) increases. As soon as this quantity of active
neurons, bn , achieves the defined value m , the regulator ceases to change the
threshold. As soon as the front edge of the sync signal appears, the pattern of