There have been multiple attempts, in the recent years, in utilizing DNN's for classification of Solar Flares [A1-A8], while some of the most powerful ML models have been tried before that [B1-B13]. Yet the question of whether we really need "Deep" NN's, or whether the "shallow" and "not-very-shallow" classical ML models would suffice is left without a proper response. While on certain tasks such as Object Detection and Speech Recognition, DNN's have undoubtedly outperformed the classical ML models, the black-box nature of deep models is indeed a limiting factor for understanding the 'why' and the 'how' of the research topics. But before we favor DNN's over classical ML models, at least in the specific task of flare classification, we would like to compare the two realms.
We are aware that a true comparison of deep and shallow models' performance, even on a specific task, is in fact not feasible, as the feature extraction process is automated in one and engineered in the other. This difference always leaves some room for the suspicion that perhaps the utilized features have not been engineered properly, and perhaps they could be further optimized or altered with other features. Having that said, this is a valid question and here we would like to try and, to a practical extend, present a semi-fair analysis to address this concern.
In this direction, this Summer Code Sprint explores some of the Machine Learning models that perhaps have not yet been utilized on multivariate time series of solar flares, despite their general success in other domains. In particular we are interested in Imaging Time Series, that is, converting time series objects into images for the purpose of classification. The main motivation for such a transformation is the tremendous success of DNN's on image data. Two of such transformations are Gramian Angular Fields (GAF) and Markov Transition Field (MTF) algorithms [Xc-Xe]. When the flare time series are transformed into image-like objects, CNN's can be utilized to classify the time series in terms of their peak flux.
We would like to compare the performance of such an approach with a more classical model, namely Time-series specific Support Vector Classifier (SVC). We hope that such a comparison would shed some light on the necessity of DNN's versus classical ML models, and potentially pave the way for the ambitious task of flare forecasting for the weather forecast community.