Method and system for tracing a learning source of an explainable artificial intelligence model

A system for tracing a learning source of an explainable artificial intelligence (AI) system, the system comprising:

• a tracing device comprising at least one processor and a computer-readable medium storing instructions that, when executed by the at least one processor, cause the at least one processor to perform operations comprising:
• receiving a desired behavior of the explainable AI model with respect to input data, wherein the desired behavior a correct classification of the input data;
• determining similarities among a plurality of learning sources with respect to the input data for the desired behavior and for a current behavior, wherein the similarities are determined using a set of filters used in a first layer of convolutional neural network (CNN);
• generating a learning graph based on the similarities among the plurality of learning sources with respect to the input data for the desired behavior and for a current behavior, wherein the current behavior comprises an erroneous classification of the input data;
• retracing a learning of the explainable AI model by iteratively comparing the learning graph for the desired behavior and for the current behavior at each of a plurality of layers of the explainable AI model starting from an outer layer wherein the learning graph is based on one or more probabilities generated for a layer through Inverse Bayesian Fusion (IBF) so as to separate learning components of one or more of a randomly selected set of learning sources;
• detecting the learning source from the randomly selected set of learning sources responsible for the current behavior based on the retracing, wherein the learning source is detected based on selecting the learning source for an output with a least distance metric in a direction of reducing gradient, and wherein the retracing is repeated until the learning source responsible for the current behavior is detected;
• correcting a classification model by unlearning with respect to the learning source responsible for the erroneous classification and learning again with respect to a corrected learning source;
• providing feedback about performance of the corrected learning source in order to check consistency of the corrected learning source;
• providing an interface for displaying the feedback about the performance to a user;
• and
• receiving an input from the user related to invoking another iteration of execution based on the feedback, thereby increasing a degree of confidence.

A non-transitory computer-readable medium storing computer-executable instructions for:

• receiving a desired behavior of the explainable AI model with respect to input data, wherein the desired behavior a correct classification of the input data;
• determining similarities among a plurality of learning sources with respect to the input data for the desired behavior and for a current behavior, wherein the similarities are determined using a set of filters used in a first layer of convolutional neural network (CNN);
• generating a learning graph based on the similarities among the plurality of learning sources with respect to the input data for the desired behavior and for a current behavior, wherein the current behavior comprises an erroneous classification of the input data;
• retracing a learning of the explainable AI model by iteratively comparing the learning graph for the desired behavior and for the current behavior at each of a plurality of layers of the explainable AI model starting from an outer layer, wherein the learning graph is based on one or more probabilities generated for a layer through Inverse Bayesian Fusion (IBF) so as to separate learning components of one or more of a randomly selected set of learning sources;
• detecting the learning source from the randomly selected set of learning sources responsible for the current behavior based on the retracing, wherein the learning source is detected based on selecting the learning source for an output with a least distance metric in a direction of reducing gradient, and wherein the retracing is repeated until the learning source responsible for the current behavior is detected;
• correcting a classification model by unlearning with respect to the learning source responsible for the erroneous classification and learning again with respect to a corrected learning source;
• providing feedback about performance of the corrected learning source in order to check consistency of the corrected learning source;
• and
• providing an interface for displaying the feedback about the performance to a user;
• receiving an input from the user related to invoking another iteration of execution based on the feedback, thereby increasing a degree of confidence.