| CPC G06N 3/0464 (2023.01) [G06N 3/0499 (2023.01)] | 20 Claims |
|
1. A computerized method, comprising:
arranging an Artificial Neural Network (ANN) in a memory of a computer circuit comprising one or more programmable processors, the ANN comprising interconnected nodes arranged into an input layer, an output layer and at least one intervening hidden layer, each node having initial parameter values stored in the memory and a non-differentiable activation function configured to generate a node output value responsive to the initial parameter values and a plurality of node input values supplied to the node;
applying input data to the input layer;
storing, in the memory, the node input values, the node output values, and a loss function value at the output layer responsive to the applied input data;
choosing a selected node;
identifying a set of downstream nodes that are interconnected along a single chain path from the selected node to the output layer;
retrieving from the memory initial parameter values associated with the selected node;
adjusting the initial parameter values of the selected node to generate adjusted parameter values;
determining a new node output value for the selected node responsive to the adjusted parameter values and the stored node input values for the selected node;
generating an updated loss function value at the output layer by determining new node update values for the set of downstream nodes based on the new node output value for the selected node using chain isolation optimization without backpropagation; and
comparing the updated loss function value to the loss function value to determine whether the adjusted parameter values generate reduced error by the ANN, and if so, replacing the initial parameter values with the adjusted parameter values in the memory for subsequent use by the selected node.
|