Classification and unsupervised clustering of LIGO data with Deep Transfer Learning

Gravitational wave detection requires a detailed understanding of the response of the LIGO and Virgo detectors to true signals in the presence of environmental and instrumental noise. Of particular interest is the study of anomalous non-Gaussian transients, such as glitches, since their occurrence rate in LIGO and Virgo data can obscure or even mimic true gravitational wave signals. Therefore, successfully identifying and excising these anomalies from gravitational wave data is of utmost importance for the detection and characterization of true signals and for the accurate computation of their significance. To facilitate this work, we present the first application of deep learning combined with transfer learning to show that knowledge from pretrained models for real-world object recognition can be transferred for classifying spectrograms of glitches. To showcase this new method, we use a data set of twenty-two classes of glitches, curated and labeled by the Gravity Spy project using data collected during LIGO’s first discovery campaign. We demonstrate that our Deep Transfer Learning method enables an optimal use of very deep convolutional neural networks for glitch classification given small and unbalanced training data sets, significantly reduces the training time, and achieves state-of-the-art accuracy above 98.8%, lowering the previous error rate by over 60%. More importantly, once trained via transfer learning on the known classes, we show that our neural networks can be truncated and used as feature extractors for unsupervised clustering to automatically group together new unknown classes of glitches and anomalous signals. This novel capability is of paramount importance to identify and remove new types of glitches which will occur as the LIGO/Virgo detectors gradually attain design sensitivity.

Figure 1

Classes of glitches in the Gravity Spy data set from the first observing run of LIGO. The No_Glitch class (periods of usual LIGO noise without transient glitches) is omitted in the figure. True GW signals from compact binary coalescence fall in the Chirp class. For each input, spectrograms with durations of 0.5 s, 1.0 s, 2.0 s, and 4.0 s are available. The objective is to predict the classes given the images.

Figure 2

Confusion matrix for Inception V3. The accuracy is 98.84% for classification on the test set. The Chirp glitch class (which is also the class that would contain true GW signals from mergers of black hole or neutron star binaries) and the Paired_Doves class (the smallest class, which only had 24 training examples) was identified with perfect precision and recall, i.e., perfect accuracy. The color in the confusion matrix only refers the values for each block.

Figure 3

Location of different classes of glitches from the test set after applying the CNN feature-extractor from our fine-tuned Inception V3 model. The t-SNE [38] algorithm was used to reduce the dimension to a 3D vector. Note that each type of glitch forms a cluster, and their relative positions depend on their morphology. Outliers may be inspected closely to verify their labels and decide whether they should belong to a new class. A new class called Reverse_Chirp was added. The CNN feature-extractor maps this class (which was not shown during training) to a unique cluster. Furthermore, this cluster is located near the Chirp class and the None_of_the_Above class, which means that the relative positions of the glitches in this feature-space is meaningful. Note that glitches in the None_of_the_Above are also grouped into smaller clusters. If points in the same cluster occur in multiple detectors at the same time, then this may indicate the presence of anomalous GW signals.