Neural Array Having Multiple Layers Stacked Therein For Deep Belief Network And Method For Operating Neural Array

A method to be performed in a neural array which is configured to form a deep neural network and implement a Restricted Boltzmann Machine, wherein the neural array has multiple selectable array values including a first array value and a second array value and is configured to cause an output signal for an array value to be an input signal for another array value, the method comprises:

• updating a first array value;
• and
• updating a second array value based on the updated first array value,
• wherein the updating the first array value includes:
• selecting the first array value;
• generating a first output signal based on a first input signal and the first array value;
• generating a reconstructed first input signal based on the first output signal and the first array value;
• generating a reconstructed first output signal based on the reconstructed first input signal and the first array value;
• and
• updating the first array value based on the first input signal, the first output signal, the reconstructed first input signal, and the reconstructed first output signal, and
• wherein the updating the second array value includes:
• generating a second input signal based on the first input signal and the updated first array value;
• and
• selecting the second array value.

A method to perform a Back Propagation Process in a neural array forming a deep neural network, the neural array has multiple selectable array values and configured to cause an output signal for an array value to be an input signal for another array value, the method comprises:

• receiving an input signal;
• performing forward propagation to generate an output signal based on the received input signal and the multiple array values;
• obtaining an error value based on the output signal and a target value for the output signal;
• performing back propagation based on the error value and the multiple array values;
• and
• updating the multiple array values based on results of the forward propagation and the back propagation.