[Pycon] [new paper] "Alessandro Re" - Architectures for the real world: retina U-nets, prototypical networks and VA-GANs in PyTorch

[info a pycon.it]

Title: Architectures for the real world: retina U-nets, prototypical networks and VA-GANs in PyTorch
Duration: 45 (includes Q&A)
Q&A Session: 0
Language: en
Type: Talk

Abstract: In this talk we will review three network architectures that we have used extensively to solve several real-world problems related to computer vision tasks in healthcare and manufacturing. We will first show how retina U-nets enable accurate and data efficient segmentation using a supervised approach, while at the same time providing object detection and classification in an end-to-end fashion. We will then delve into prototypical networks, which were originally created for classification in few-shots learning contexts, and show how they can be effective at learning from limited datasets, refining datasets, communicating outputs to end-users and enabling them to have some control on decision boundaries. Last, we will show how Visual Attribution GANs can provide dense identification of anomalies under weak supervision. We will review the implementation of all these architectures in PyTorch and see them in the wild on a number of practical use cases.

Tags: [u'machine-learning', u'deep learning', u'Pytorch', u'neural network']