PyData Berlin November meetup

Our Novermber meetup is kindly hosted by Wayfair on Kopenicker Strasse and will have a strong neural networks flavour to it. Doors open at 18:45, drinks and food is kindly sponsored by Wayfair.

We have two speakers Stefan Otte and Johannes Rieke.

Johannes Rieke (~30 minutes): Opening the black box: Interpreting deep convolutional networks in PyTorch

Convolutional neural networks (CNNs) are powerful machine learning models in computer vision. Despite their intriguing performance, CNNs are hard to interpret. In this talk, I will present several methods that explain the classification decisions of a CNN by highlighting relevant regions in the input image. I will show results on 2D images as well as 3D brain scans (for automated detection of Alzheimer's disease; see my recent paper: https://arxiv.org/abs/1808.02874). After the talk, you will be able to interpret your own PyTorch networks, using the code I published at https://github.com/jrieke/cnn-interpretability.

Johannes is a Masters Student in Computational Neuroscience (TU Berlin). He has worked on deep learning for the past two years, sometimes just for himself, sometimes for SAP, Charite Berlin and University of Amsterdam. He has sticked to Python for the last five years (and counting!), most recently fell in love with pytorch.

You can follow Johannes at @jrieke (https://twitter.com/jrieke)

************

Stefan Otte (~30 minutes) Fun with content-based image retrieval (CBIR) with neural networks

"Content-based image retrieval (CBIR), [...] is the application of computer vision techniques to the image retrieval problem, that is, the problem of searching for digital images in large databases [...]." https://en.wikipedia.org/wiki/Content-based_image_retrieval

CBIR is a widely used technique that rapidly gaining in popularity. With it you can search for cute dog pictures, find similar products or art , and even more. Up until recently CBIR used "classic computer vision" algorithms, but neural networks are gaining traction.

This talk consists of two parts:
1. A *brief introduction to CBIR with neural networks*: the general setup and architecture, feature representations, similarity measures, metrics and losses, datasets, different ways of training, common problems, and differences to classic computer vision algorithms.

2. *Lessons learned*: hands-on implementation of CBIR with PyTorch, speeding up the search, and some (fun) ideas of what you can do with this setup.

Stefan is a machine learner, data scientist, ex-roboticist, and consultant. He likes turning crazy ideas into production systems using data science and software engineering. You can find him at https://sotte.github.io/