US 11,922,296 B2
Electrical networks using analytic loss gradients for design, analysis and machine learning
Jack David Kendall, Millbrae, CA (US)
Assigned to Rain Neuromorphics Inc., San Francisco, CA (US)
Filed by Rain Neuromorphics Inc., San Francisco, CA (US)
Filed on Jul. 27, 2022, as Appl. No. 17/875,223.
Claims priority of provisional application 63/243,998, filed on Sep. 14, 2021.
Claims priority of provisional application 63/226,637, filed on Jul. 28, 2021.
Prior Publication US 2023/0040533 A1, Feb. 9, 2023
Int. Cl. G06N 3/063 (2023.01)
CPC G06N 3/063 (2013.01) 18 Claims
OG exemplary drawing
 
1. A system, comprising:
a plurality of inputs;
a plurality of outputs;
a plurality of hardware neurons;
a plurality of nodes between the plurality of inputs and the plurality of outputs, the plurality of nodes including a plurality of hidden nodes, a portion of the plurality of hidden nodes being coupled to the plurality of hardware neurons; and
a plurality of weights including a plurality of programmable impedances coupled between at least a portion of the plurality of nodes, the plurality of weights being coupled to connections, the connections being between the plurality of weights and the portion of the plurality of hidden nodes;
wherein the connections between the plurality of nodes are configured such that a gradient of a loss function for the system is explicitly computable using symmetric solution submatrices, wherein the symmetric solution submatrices are part of a modified nodal analysis (MNA) solution matrix, the symmetric solution submatrices describing the connections and corresponding to the connections being reciprocal with respect to the plurality of outputs; and
wherein the plurality of weights are programmable based on the gradient by perturbing a plurality of states of the plurality of outputs, measuring a plurality of resulting perturbations in an electrical characteristic of the connections, and programming the plurality of programmable impedances based on the plurality of resulting perturbations in the electrical characteristic for the connections;
wherein the measuring the resulting perturbations in the electrical characteristic for the connections corresponds to measuring the gradient for weight of the plurality of weights.