Low rank adaptation in multiclass deep learning classifiers

Abstract

The computer-implemented method utilizes machine learning models to classify input data records. The method includes selecting machine learning models from a plurality of machine learning models using a router model and generating classification inferences for each selected model. The classification inferences include factors indicating the probability of the input data record belonging to various candidate classes. Any one of the plurality of machine learning models is configured to use a weight matrix calculated as the sum of a base weight matrix and a model-associated difference matrix obtained as a product of two low-rank matrices. The classification inferences from selected models are then combined to create an overall classification inference, providing factors related to the probability of the input data belonging to each candidate class. This method enables effective classification by leveraging multiple machine learning models.