Stars_detection_with_LSST_time_series.pdf

Detecting objects in deep space using deep learning

Transcript

1. Objecte detection with LSST time series Martin Eriksen and Skybert Svendson Image credit: LSST Corporation
2. Next generation telescope - 8.4 meter telescope. - 10 years of observations. - Observations in 6 optical bands. - 200 000 images per year. - 37 billion stars and galaxies - Accurate measurement of dark matter and dark energy. - 20TB of data per night. - Require 100PB of storage.
3. The data - Simulations of LSST data for first 3 year of operations. - Provides time series in 6 bands and additional metadata [e.g. distance] - 14 classes. Unknown interpretation. - 7848 training objects, 3.5 million tests objects. - Part of the PLAsTiCC Kaggle challenge. Easy features!
4. Data preprocessing - Creating features from time series. - Natively scale Python. - Allows running preprocessing of the test set, which is later than memory. - Automatically use all CPU cores. import dask.dataframe as dd df = dd.read_csv(path_in) df = df.repartition(npartitions=1000) df.to_parquet(path_out)
5. Network architecture - Experimented with different unsupervised methods. - Used the PyTorch framework: - On the right is one example of a neural network.
6. Results - Quite a simple network gives 72% accuracy. - Large class imbalanced. Optimize using a custom loss function. - Some classes trickier than others.
7. Conclusions - The Large-aperture Synoptic Survey Telescope (LSST) will be very important for understanding dark energy, dark matter, detecting super novae and transients. - DASK. Allows for parallel out-of-core processing from Python. - PyTorch. Flexible framework for deep learning. - Irregular sparse time series. Annoyingly tricky. - AI Saturday. Why not join?