Symposium F.MT02: Multimodal, Functional and Smart Scanning Probe Microscopies for Characterization and Fabrication
Joshua Agar, Lehigh University
Practical Deep Learning in Scanning Probe Microscopy
Written by Emma Perry
Whilst open source machine learning algorithms promise speed and low mean square values, Joshua Agar offers a word of caution to scientists planning to use them without carefully studying the mathematics and applying a rigorous testing regime.
The analysis of the band excitation piezoresponse spectroscopy is a problem that could greatly benefit from machine learning. These experiments produce a complex 4 channel hyperspectral image.
In this talk a dictionary learning method is compared to a neural network deep autoencoder method. First the methods are applied to a toy data set and then a benchmark band excitation piezoresponse spectroscopy dataset. It is quickly seen that the dictionary learning method is the simplest and fastest. But this method fits based on the mean square error and the subtleties, that carry physical meaning, are lost. The neural network deep autoencoder method does manage to deliver the detail of the functions. In the live Q & A session Joshua Agar warns that there is no one solution for all. In your own work you must start with the simplest model, rigorously test it, and work from there.