US 12,456,054 B2
Gradient descent training for defensible artificial intelligence
Jeremy Straub, Fargo, ND (US)
Assigned to NDSU Research Foundation, Fargo, ND (US)
Filed by NDSU Research Foundation, Fargo, ND (US)
Filed on Mar. 7, 2022, as Appl. No. 17/688,659.
Prior Publication US 2023/0281452 A1, Sep. 7, 2023
Int. Cl. G06N 3/082 (2023.01); G06N 3/048 (2023.01)
CPC G06N 3/082 (2013.01) [G06N 3/048 (2023.01)] 14 Claims
OG exemplary drawing
 
1. A system for defensible artificial intelligence networks, the system comprising:
processing circuitry; and
one or more memory devices including instructions, which when executed by the processing circuitry, configure the processing circuitry to:
obtain a plurality of input facts from the one or more memory devices;
obtain a plurality of target network output results; and
train an expert system rule network by iteratively performing a plurality of steps of:
generating a training rule network;
generating a plurality of training output results based on the training rule network and the plurality of input facts;
generating a revised rule network based on a comparison between the plurality of training output results and the plurality of target network output results, the revised rule network including a plurality of network rules, each of the plurality of network rules having a plurality of associated fact weightings, each weighting identifying a mapping of the plurality of input facts to an output fact;
identifying a first network rule within the plurality of network rules, the first network rule associated with a rule effect on the expert system rule network;
determining the rule effect crosses a rule pruning threshold;
removing the first network rule responsive to determining the rule effect crosses the rule pruning threshold;
generating a plurality of trained network outputs based on the expert system rule network;
generating a test network by suspending the first network rule within the expert system rule network;
generating a plurality of test network outputs by running the test network without the first network rule;
determining a performance metric based on a comparison between the plurality of trained network outputs and the plurality of test network outputs, wherein determining the rule effect crosses the rule pruning threshold includes determining the performance metric satisfies a network performance threshold; and
determining, responsive to removing the first network rule, that all rules have been removed from the expert system rule network.