Hardware for Neural Networks - Neural Networks and Micromechanics - page 130

register, and Y is the code stored in RAM; & is conjunction; and is negation. This

operation continues until, in code Z , a quantity of unit elements becomes less than
m . The resulting vector is called the normalized input vector of the neural network.
The process of input information coding is terminated.
If the neural network was not trained previously, then the training mode begins
when the synaptic weights of interneuronal connections are reset to 0. After this, the
sets of parameters that must be associative, and connected with each other, are
supplied to the input of the neural network. Each of the input sets of the parameters is
coded, and the code obtained in the shift register is maintained in RAM. Next, the
process of training occurs, though we will not discuss this in detail. Note that the
numbers of the synaptic matrix rows corresponding to active neurons are extracted
according to the numbers of active neurons, and these numbers are converted into the
addresses of RAM. Each obtained matrix row M i undergoes the following operation:

M i ¼ M i U Z ð & P Þ; (7.1)

where P is the vector formed with the aid of the pseudorandom number generator.
The obtained vector M i * is written in RAM with the same address where the vector
M i was stored. The experiments with this algorithm of training showed that, in the
neural networks, structures are formed that correspond to the neural ensembles.
In the mode of associative recall of information, sets of input parameters in
which one part coincides with one of the ensembles and another part is selected
randomly as a noise component are supplied to the input. The task of the neural
network is to restore the missing parameters from the noise set and to eliminate the
noise component. Coding of the input set of parameters in this mode is accom-
plished exactly as in the mode of training. Associative restoration is achieved in the
process of cyclic recalculation of neuron activity.
In the mode of decoding, the neural network must decode to what class the
excited neuron ensemble corresponds, which is achieved as follows. The obtained
binary vector of neuron activity is compared with all vector codes of the classes
stored in RAM. If a quantity of coinciding elements exceeds a certain threshold, it is
assumed that the code of this class is present in the resulting output vector of the
neural network.
This neurocomputer was used for texture recognition as described earlier.

7.2 Neurocomputer B-512

As a result of work with the neurocomputer described above, the need arose to
create a more productive neurocomputer with greater memory and an improved
system of instructions. To explore the possibilities of the new version of the
neurocomputer, the authors of this topic developed a neurocomputer emulator,
and later, the neurocomputer was developed with the participation of D. A.