US 12,242,970 B2
Incremental cluster validity index-based offline clustering for machine learning
Leonardo Enzo Brito Da Silva, Natal (BR); Donald C. Wunsch, II, Rollo, MO (US); and Nagasharath Rayapati, San Ramon, CA (US)
Assigned to Guise AI, Inc., Rolla, MO (US)
Filed by Guise AI, Inc., Rolla, MO (US)
Filed on Aug. 16, 2021, as Appl. No. 17/402,938.
Claims priority of provisional application 63/066,209, filed on Aug. 15, 2020.
Prior Publication US 2022/0318632 A1, Oct. 6, 2022
Int. Cl. G06N 3/082 (2023.01); G06N 3/04 (2023.01); G06N 3/043 (2023.01)
CPC G06N 3/082 (2013.01) [G06N 3/0409 (2013.01); G06N 3/043 (2023.01)] 18 Claims
OG exemplary drawing
 
1. A method for machine learning, comprising:
a) configuring an ART A (Adaptive Resonance Theory A) module of a fuzzy ARTMAP (Adaptive Resonance Theory Predictive Mapping) neural network as a fuzzy ART module;
b) configuring an offline incremental cluster validity index (iCVI) module as a second input module to the fuzzy ARTMAP neural network;
c) initializing a data set partition;
d) initializing the ART A module, the offline iCVI module, and a map field of the fuzzy ARTMAP neural network corresponding to the initialized data set partition, said ART A module further having an ART A vigilance parameter;
e) inputting pre-processed versions of a data set into both the ART A module and the offline iCVI module;
f) in the offline iCVI module, computing temporary iCVI values for an assignment of a current sample of the pre-processed versions of the data set to each of a plurality of clusters in a current data partition and generating a current sample label for the current sample as a function of the computed temporary iCVI values;
g) applying the current sample label to a vigilance test of the map field, said map field having a map field vigilance parameter;
h) assigning a sample to an ART A category and an associated cluster mapped via the map field when both an ART A and the map field vigilance tests are simultaneously satisfied, but when the ART A vigilance parameter is satisfied but the map field vigilance parameter is not satisfied, then causing the ART A module to change its vigilance parameter;
i) incrementally updating an ART A category weight vector and a corresponding map field weight vector when both the ART A and the map field vigilance tests are simultaneously satisfied;
j) creating a new ART A category weight vector and a new corresponding map field weight vector when no existing ART A category simultaneously satisfies the ART A and the map field vigilance tests;
k) incrementally updating iCVI variables and validation measures when assignments of samples of the data set to respective clusters change and adjusting assignments of the ART A categories to clusters according to the map field, thereby adjusting input-output relationship of the neural network; and
l) generating subsequent data partitions with a multi-prototype cluster representation via a categories-to-clusters mapping of the map field.