US 12,353,994 B2
System and method for reasoning about the diversity and robustness of an ensemble of classifiers
Shantanu Rane, Menlo Park, CA (US); Alejandro E. Brito, Mountain View, CA (US); and Hamed Soroush, San Jose, CA (US)
Assigned to Xerox Corporation, Norwalk, CT (US)
Filed by Palo Alto Research Center Incorporated, Palo Alto, CA (US)
Filed on Jan. 26, 2021, as Appl. No. 17/158,631.
Prior Publication US 2022/0237443 A1, Jul. 28, 2022
Int. Cl. G06N 3/08 (2023.01); G06F 18/214 (2023.01); G06F 18/2415 (2023.01); G06N 20/10 (2019.01); G06N 20/20 (2019.01)
CPC G06N 3/08 (2013.01) [G06F 18/214 (2023.01); G06F 18/2415 (2023.01); G06N 20/10 (2019.01); G06N 20/20 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A computer-executable method for facilitating reasoning about classifiers, the method comprising:
determining a plurality of neural networks in a machine learning environment;
receiving, by a respective neural network, an original input image;
processing, by convolutional layers of the respective neural network, the original input image;
obtaining, based on the processing by the convolutional layers, features associated with the original input image;
passing the features to an input layer and a hidden layer of the respective neural network, wherein the hidden layer comprises a penultimate layer of the respective neural network;
obtaining, based on the features passed to the penultimate layer, an output of the penultimate layer comprising activations which indicate an intermediate representation of the original input image to the respective neural network;
deriving, from the respective neural network, a linear model based on the activations of the penultimate layer of the respective neural network, wherein the activations are associated with fewer parameters than the original input data;
training the linear model based on the activations of the penultimate layer;
mapping parameters of the trained linear model into a version space in the machine learning environment;
obtaining an ensemble of classifiers by deriving, from each of the plurality of neural networks, a collection of linear models each belonging to a version space in the machine learning environment;
measuring a robustness of the ensemble of classifiers based on removing a version space boundary from a respective version space;
displaying, on a display screen of a computing device associated with a user, information associated with the ensemble of classifiers;
in response to determining a command performed by the user, updating the ensemble of classifiers, including measuring the robustness of the updated ensemble of classifiers; and
in response to updating the ensemble of classifiers, displaying, on the display screen, updated information associated with the updated ensemble of classifiers.