Efficient few-shot machine learning for classification of EBSD patterns

Kevin Kaufmann,Xiao Liu,Hobson Lane,Kenneth S Vecchio

doi:10.1038/s41598-021-87557-5

Kevin Kaufmann, Xiao Liu + Show 2 more

Open Access

https://doi.org/10.1038/s41598-021-87557-5

Copy DOI

Abstract

Deep learning is quickly becoming a standard approach to solving a range of materials science objectives, particularly in the field of computer vision. However, labeled datasets large enough to train neural networks from scratch can be challenging to collect. One approach to accelerating the training of deep learning models such as convolutional neural networks is the transfer of weights from models trained on unrelated image classification problems, commonly referred to as transfer learning. The powerful feature extractors learned previously can potentially be fine-tuned for a new classification problem without hindering performance. Transfer learning can also improve the results of training a model using a small amount of data, known as few-shot learning. Herein, we test the effectiveness of a few-shot transfer learning approach for the classification of electron backscatter diffraction (EBSD) pattern images to six space groups within the left( {4/m overline {3} 2/m} right) point group. Training history and performance metrics are compared with a model of the same architecture trained from scratch. In an effort to make this approach more explainable, visualization of filters, activation maps, and Shapley values are utilized to provide insight into the model’s operations. The applicability to real-world phase identification and differentiation is demonstrated using dual phase materials that are challenging to analyze with traditional methods.

Highlights

Deep learning is quickly becoming a standard approach to solving a range of materials science objectives, in the field of computer vision
Deep learning algorithms are of significant interest owing to their excellent performance without significant feature engineering, and the ubiquity of these methods will likely continue owing to the outperformance of systems directly designed by humans
Two general strategies exist for training convolutional neural networks: (1) the weights can be randomly initialized, or (2) the weights can be transferred from a model pre-trained on a separate but related task, often in a nearby domain with significantly more data, and refined for the current objective

Summary

Introduction

Deep learning is quickly becoming a standard approach to solving a range of materials science objectives, in the field of computer vision. While often difficult to assess how and why these ‘black box algorithms’ are capable of performing these tasks, these methods can provide significant value or spark new insights[14,15] Application of these tools to image-based tasks in materials science has proved to be useful for c lassification16–19, segmentation[20,21,22], and other objectives[23,24,25]. In deep learning and computer vision, learning visual models of object categories has notoriously required tens of thousands of training examples[36]; recent research has demonstrated that it is possible to classify images accurately using relatively few labeled examples with the appropriate combination of pretraining of the CNN layers on unrelated image classification training s ets[37,38], adversarial or unsupervised learning[39,40], network p runing[41], and micro architecture tuning[42]. The increased efficiency allows for a lesser number of images to be used in training, referred to as a “few shots”

Objectives

Methods

Results