Improving deep neural networks via prototype factorization

Abstract

The deep neural network is improved via prototype factorization. A method may include receiving a set of images, analyzing the images, selecting an internal layer, extracting neuronal activations, factorizing the neuronal activations via a matrix factorization algorithm to select prototypes and generate weights for each of the selected prototypes, replacing the neuronal activations of the internal layer with the selected prototypes and the weights of the selected prototypes, and generating a plurality of neuronal activations of the internal layer. Receiving a second image set, classifying the second image set using the prototype and the weight, displaying the second image set, the selected prototype and the weight, displaying a prediction result and a reference truth value of the second image set, and providing an error image based on the prediction result and the reference truth value; identifying an erroneous prototype of the selected prototype associated with the erroneous image; the error weights of the error prototype are ranked, and a new image class is output based on the error prototype being one of the error weights ranked at the top.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS [0002] This application claims the benefit of US Provisional Application No. 63 / 108,192, filed October 30, 2020, the entire disclosure of which is incorporated herein by reference. technical field [0003] The present disclosure generally relates to systems and methods for image classification, and operations based on the resulting classifications. Background technique [0004] A typical deep neural network is a complex black-box model, and its decision-making process is incomprehensible even for experienced machine learning (ML) practitioners. Therefore, despite state-of-the-art performance in many challenging ML tasks, their use in mission-critical scenarios may be limited. Moreover, in recent years, deep neural networks (DNNs) have been increasingly used in various application domains due to their advanced performance in many challenging machine learning tasks. However, their lack of interpretability can lead to trustwort...

AI Technical Summary

Problems solved by technology

However, their lack of interpretability can lead to trustworthiness and fairness issues, and also makes model diagnosis a difficult task.

Images