US 12,423,526 B1
Intelligent detection of bias within an artificial intelligence model
Ali Fathi, San Francisco, CA (US); and Samuel A. Assefa, Dallas, TX (US)
Assigned to U.S. Bank, Minneapolis, MN (US)
Filed by U.S. Bank, Minneapolis, MN (US)
Filed on Jun. 2, 2025, as Appl. No. 19/225,268.
Application 19/225,268 is a continuation of application No. 19/062,986, filed on Feb. 25, 2025, granted, now 12,367,349.
Application 19/062,986 is a continuation of application No. 18/421,318, filed on Jan. 24, 2024.
Int. Cl. G06F 17/00 (2019.01); G06F 40/284 (2020.01); G06F 40/30 (2020.01); G06F 40/40 (2020.01)
CPC G06F 40/30 (2020.01) [G06F 40/284 (2020.01); G06F 40/40 (2020.01)] 20 Claims
OG exemplary drawing
 
1. A method for producing an embedding model, the method comprising:
receiving identification of a class;
gathering or generating a plurality of pairs of words or tokens associated with the class;
calculating a bias direction of the class in an embedding space, wherein calculating the bias direction includes:
obtaining a corpus of token pairs related to the class;
calculating token embeddings for each token of the token pairs of the corpus using the embedding model;
performing principal component analysis on the embeddings such that variance associated with the class comes to lie on a first principal component that expresses a biased portion of a meaning of the tokens; and
determining the bias direction using the first principal component;
for respective tokens in a series of tokens, calculating a protected gradient score with respect to the bias direction;
aggregating protected gradient scores of the respective tokens to form a series-level protected gradient score for the series of tokens;
determining a fairness indicator based on the protected gradient score for the series; and
responsive to the fairness indicator being above a threshold, permitting deployment of the embedding model.