| Abstract : | The time complexity of matched-filtering searches directly scales with the number of templates and their duration. As the interferometers approach their design sensitivity, along with more detectors joining this venture and with the advent of new models to probe into a more expansive parameter space, the time complexity of these searches is bound to increase tremendously. Deep learning is a possible way to make these searches low latency by leveraging the fact that the model has to be trained only once and can be deployed online faster than the current search algorithms. The search for Intermediate Mass Black Hole (IMBH) binaries is even more challenging, owing to their short duration and similarity in morphology with frequently occurring noisy glitches. We have developed a Deep Transfer Learning model with Inception-v3 architecture, which extracts features from the spectrograms generated using continuous wavelet transform, fixes the convolutional base and fine-tunes the dense layers for prediction. This model classifies the dataset into different types of glitches and signals from IMBH binaries. We used non-spinning, quasicircular IMBH binary injections for generating the dataset. The model achieves a training accuracy of 96.27% and a testing accuracy of 90.77%, with 99.25% accuracy in detecting IMBH binaries. We tested this model on massive binaries in O3 data, where these events were detected confidently. This talk presents the detailed analysis framework, including the network architecture, datasets, model development, and tests on massive O3 binaries. |
