Neural network with weight-controlling auxiliary path

A circuit for use in a neural network, comprising:

• a first plurality of neurons in a main path, each of the first plurality of neurons configured to receive a plurality of input signals through a first plurality of weighting elements and produce a first plurality of sum-of-products signals by weighting each of the input signals and adding the weighted input signals together, each of the first plurality of weighting elements connected to or disconnected from each of the first plurality of neurons by a first plurality of switches;
• a second plurality of neurons in the main path, each of the second plurality of neurons configured to receive some or all of the first plurality of sum-of-products signals through a second plurality of weighting elements and produce a second sum-of-products signal by weighting each of the first plurality of sum-of-products signals and adding the weighted sum-of-products signals together, each of the second plurality of weighting elements connected to or disconnected from the each of the second plurality of neurons by a second plurality of switches;
• and
• a third plurality of neurons in an auxiliary path, each of the third plurality of neurons configured to receive the plurality of input signals through a third plurality of weighting elements, produce a third sum-of-products signal by weighting each of the input signals and adding the weighted input signals together, and control positions of the first or second plurality of switches based on the third sum-of-products signal.

A method of operating a neural network, comprising:

• receiving in a first neuron of a first plurality of neurons in a main path a plurality of input signals through a first plurality of input weighting elements;
• receiving in a second neuron of a second plurality of neurons in an auxiliary path the plurality of input signals through a second plurality of input weighting elements;
• summing in the second neuron the weighted input signals from the second plurality of input weighting elements;
• operating a plurality of switches based upon the sum from the second neuron to selectively connect the first plurality of input weighting elements to the first neuron;
• and
• summing in the first neuron the weighted input signals from the first plurality of input weighting elements.

A method of operating a neural network, comprising:

• receiving, in a first layer of a main path comprising a plurality of neurons, a plurality of input signals, each neuron in the first layer receiving some of the input signals through one of a first plurality of input weighting networks;
• receiving, in a second layer of the main path comprising a plurality of neurons, a plurality of signals from the first layer of neurons, each neuron in the second layer receiving some of the signals from the first layer of neurons through one of a second plurality of input weighting networks;
• receiving in a third plurality of neurons in an auxiliary path the plurality of input signals through a third plurality of input weighting elements;
• summing in the third neuron the weighted input signals from the third plurality of input weighting elements;
• operating a plurality of switches using the sums from the third plurality of neurons to selectively connect the first plurality of input weighting networks to the neurons in the first layer or the second plurality of input weighting networks to the neurons in the second layer;
• summing in the neurons in the first layer the weighted input signals from the first plurality of input weighting elements;
• and
• summing in the neurons in the second layer the signals from the first layer of neurons from the second plurality of input weighting elements.