Nano Lett. 17, 3113–3118 (2017)

Dimensionality reduction is a fundamental aspect of machine learning algorithms. It involves the approximation of high-dimensional multivariate input data as a compact combination of few fundamental elements in a process known as feature extraction, leading to more-efficient information storage and analysis. Feature extraction has wide implications for predictive modelling based on big data sets — however, it is computationally intensive.

Now, Choi et al. report on the efficient exploitation of neuromorphic hardware architectures to perform principal component analysis — a linear version of the feature extraction algorithm. The researchers fabricate a 9 × 2 crossbar array of memristors based on 10-nm-thick Ta2O5 switching layers sandwiched between NiCr/Pd and Ta/Pd electrodes. These memristors are analogue and their resistance state can be adjusted incrementally by the application of voltage pulses affecting the oxygen vacancies profile in Ta2O5.

Each of the nine rows of the memristor array is fed with an input voltage pulse. Encoded in the duration of each pulse is one specific mass property of a cell and, in an initial training process based on a breast cancer cell database, the researchers use 100 nine-input sets — relative to 50 benign and 50 malignant cells. The trained network then allows the researchers to analyse 600 unlabelled data sets. A clustering analysis based on the learned features shows that the unsupervised learning process in the array led to successful classification of benign and malignant cells in 97% of cases.