Abstract: Time series classification is the problem of assigning classes/labels to samples that have one or more time series as features, e.g., in order to identify gestures or activities based on motion sensor data, or to classify land use based on periodic satellite imagery. The state of the art in this space in terms of overall classification accuracy is an ensemble classifier, called HIVE-COTE 2.0, that is extremely computationally expensive (making it slow to train and impractical for large datasets). I will detail some research on an alternative technology that we developed, called MINIROCKET, that is based on linear combinations of pooled convolutional filters constructed from permutations of just two distinct filter weights. This results in a classifier that achieves near-state-of-the-art classification performance while being several orders of magnitude faster than HIVE-COTE 2.0 or competing deep neural network models. A complete run (training and inference) of MINIROCKET on a standard set of datasets from the UCR time series repository takes in the order of minutes, versus two weeks for HIVE-COTE 2.0.

One of the reasons MINIROCKET is so fast is due to its use of a classifier based around standard squared-error ridge regression. While fast, one drawback of this approach is that it is unable to produce probabilistic predictions. I will demonstrate some very recent work on how to use regular ridge-regression to train L2-regularized multinomial logistic regression models for very large numbers of features, including choosing a suitable degree of regularization, with a time complexity that is no greater than single ordinary least-squares fit. This allows for models based on the MINIROCKET technology to provide well calibrated probabilistic predictions with minimal additional computational overhead.

Speaker:  Daniel Schmidt is an Associate Professor of Computer Science at the Department of Data Science and AI, Monash University, Australia. He obtained his PhD in the area of information theoretic statistical inference in 2008 from Monash University. From 2009 to 2018 he was employed at the University of Melbourne, working in the field of statistical genomics (GWAS, epigenetics and cancer). Since 2018 he has been employed at Monash University in a teaching and research position. His research interests are primarily time series classification and forecasting, particularly at scale, and Bayesian inference and MCMC, with an emphasis on sparsity and shrinkage, Bayesian optimisation and Bayesian function approximation. He also has a keen interest in the best ways to provide statistical/machine learning education.

Affiliation: Monash University

Place of Seminar:  Kumpula Exactum CK111 (in person) & zoom ( Meeting ID: 640 5738 7231 ; Passcode: 825217)