AI3SD Video: When charge transport data are a worm – a transfer learning approach for unsupervised data classification

Abstract

Advanced data analysis methodologies, and in particular dimensionality reduction techniques, are now used more and more widely in the single-molecule charge transport community. They allow for comprehensive exploration of large datasets, where data display significant variance and sometimes contain (unknown) sub-populations. To this end, unsupervised approaches, which do not rely on class labels or pre-defined expectations can be advantageous. Multi-Parameter Vector Classification (MPVC) is one example and PCA-based methods have also been employed in this context [1,2,3]. We have recently shown how Transfer Learning may be employed to identify and quantify hidden features in single-molecule charge transport data [3]. Using open-access neural networks such as AlexNet, trained on millions of seemingly unrelated image data, feature recognition then does not require network training with application-specific data. Instead, the network recognises features in the input that it had learned in other contexts and, for example, identifies different shapes in conductance-distance traces as images of different worm species. Thus, our results show how Deep Learning methodologies can readily be employed for unsupervised data classification, even if the amount of problem-specific, ‘own’ data is limited. [1] M Lemmer, MS Inkpen, K Kornysheva, NJ Long, T Albrecht, “Unsupervised vector-based classification of single-molecule charge transport data”, Nat. Comm. 2016, 7, 12922. [2] T Albrecht, G Slabaugh, E Alonso, SMMR Al-Arif, “Deep learning for single-molecule science”, Nanotechnology 2017, 28 (42), 423001. [3] A Vladyka, T Albrecht, “Unsupervised classification of single-molecule data with autoencoders and transfer learning”, Mach. Learn.: Sci. Technol. 2020, 1, 035013.