Keywords: atomic force microscopy, machine learning, cell phenotype, image recognition, reduction of data space
Summary:Recent large interest in machine learning (ML) analysis has been extended to atomic force microscopy (AFM). In this talk, I will address three questions that typically arise when AFMers start applying ML to classify the AFM images: 1) What mode is the best to use? 2) How many images is sufficient to run machine learning classifiers and believe the results? 3) How do I know that the obtained classification is not an artifact of some fine-tuning? In this talk, I will suggest answers to the questions. In short, the mode of operation should be robust and repeatable. This is the key. The number of images required for classification can be roughly estimated using “the rule of 10”. A more precise algorithm to identify the required number of images will be described. To be confident that one doesn’t deal with an artifact of fine-tuning or overtraining a classifier, we suggest a rather simple yet powerful method of estimating the exact amount of overtraining by using randomization of the initial classification. Examples will be given.