This is a meetup where focus is on cutting edge usage of machine learning and thus the entry barrier will be a little demanding. On the other hand we will do our best to give excellent input on the topics selected.
Learning Neural Networks to Solve Reasoning Tasks
Deep neural networks are very good at recognizing objects, but when it comes to reasoning about their interactions even state of the art neural networks struggle. In this talk, Rasmus will introduce a neural network module that model objects and interactions naturally and show how we can use it for solving tasks requiring complex relational reasoning such as Sudokus.
• Msc in medical engineering from Technical University of Denmark
• Created a popular (now deprecated) Matlab toolbox for deep learning.
• Worked at Tradeshift for the last 6 years.
• Created the machine learning team there.
• Finished PhD in deep learning for information extraction.
Leveraging the Exact Likelihood of Deep Latent Variable Models
Deep latent variable models (DLVMs) combine the approximation abilities of deep neural networks and the statistical foundations of generative models. In this talk, I will consider two inferential problems for DLVM: estimation and missing data imputation. First, I will discuss maximum likelihood estimation for DLVMs and show that the most unconstrained models used for continuous data have an unbounded likelihood function. Secondly, I will present a new algorithm for missing data imputation using the exact conditional likelihood of a DLVM. I will show that our algorithm consistently and significantly outperforms the usual imputation scheme used for DLVMs.
Jes Frellsen is an Associate Professor at the IT University of Copenhagen (ITU). His main area of research is Machine Learning with an emphasis on Deep Learning, Bayesian modelling and inference, and applications in Bioinformatics. Before joining the ITU, he was a postdoctoral researcher with Professor Zoubin Ghahramani in the Machine Learning group at the University of Cambridge. Prior to that, he was a graduate student and postdoc at the Bioinformatics Centre, University of Copenhagen, where he worked with Professor Anders Krogh and Associate Professor Thomas Hamelryck.
Turbine Blade Defect Detection: Going from zero to product in a year
This talk will focus on how industry problems, here wind turbine blade defect detection, can be solved using machine learning and especially deep convolutional networks. I'll dive into the process of going from scratch to a fully operational service and product in less than a year. I'll talk about the systems and requirements around the AI heart of any service and how we handle scaling. Finally, I'll share some of the learnings we have made until now and the challenges that we are facing.
• Master and Ph.D in bio-informatics Copenhagen University
• Post-doctoral work in bio-statistics Princeton University
• 7 years of research and management experience in LEO-Pharma
• Part of founding team for LEO-Pharmas innovation unit in Boston, MA
• Founding partner in Wind Power Lab
• Working in the interface between new technologies, machine learning and biology with pioneering groups from MIT, MGH, and various start-ups
Food: Drinks and food will be provided in the end of the event.
Organizer: Johan Bender, IDA AI, Nordic AI