Humberto Farias holds a BA and a Master's in Computer Science from Universidad Técnica Federico Santa María (UTFSM). He has developed most of his work within the nascent field of astroinformatics, which seeks to provide solutions for the computational challenges faced by astronomy. He is part of the team in charge of the Chilean Virtual Observatory (ChiVO) and is finishing a PhD at the UTFSM, focusing his research on the storage (tensor methods) and processing (DeepLearning) of multidimensional data cubes obtained from multi band observations. Humberto has worked for the Big Data Project of the European Southern Observatory (ESO) and other institutions. He is also a reviewer of projects on Big Data and machine learning for Chilean government.
From Semantic segmentation to Instance Segmentation using DeepLearning.
The tasks of location, classification, and segmentation are known and applied by astronomers in various problems such as: Morphological classification of galaxies, Transient detection, search for supernovae among others.
Is widely known in this decade will see a series of astronomical mega-projects coming into operation producing complex data whose dimensionality and volume will exceed any current scale. This requires the application of a new generation of machine learning (Deeplearning) models for classification, location, and segmentation.
In this tutorial we will cover the latest advances in Deep learning applied to Semantic segmentation, Object localization and Instance Segmentation. The tutorial modality will be divided into blocks of 30 minutes as follows:
Part 1: Introduction (Theory)
Part 2: Semantic segmentation (U-NET)
Part 3: Object localization (YOLO3)
Part 4: Instance Segmentation (Mask R-CNN)
Prerequisites: Intermediate Python knowledge is strongly recommended.
Level: We assume you are comfortable with deep learning basics as layers, neuron, activation function, loss function among other basics concepts.
Tensor Mask R-CNN: A Tensor-based Deep Learning approach for fast morphological classification and segmentation of Galaxies.
Modern astronomy has evolved into a highly computational scenario, as a result of the volume and complexity of the data that must be processed and analyzed. This has motivated the application of advanced computer vision techniques such as Deep Learning, which have shown outstanding results in several applications. However, the implementation of these models usually lack a key factor that impedes the incorporation of such models into complex scientific problems that demand large processing times: consideration of speed of inference as a constraint of the problem. To solve this problem, several authors have presented proposals that incorporate tensor layers into the architecture of CNN networks, showing satisfactory results. Following this rationale, we propose Tensor Mask R-CNN, a combination of a state-of-the-art regional convolutional neural network and tensor methods. This novel approach is based on the incorporation of Tensor Layers (T-L), that is, applying tensor decomposition over fully connected layers in the backbone of a Mask R-CNN network. These tensor layers provide more powerful and cheaper parameterizations for fully connected layers. This results in compressed parameter spaces that lead to a reduction in computational costs and, consequently, a reduction in inference time. As a test of our proposal, we apply Tensor Mask R-CNN to the classification of galaxies based on their morphology, where we validate the increase in inference speed, and reduced costs therefore, against the same architecture without tensor layers.