2019 | Tech

10.15-10.45: SCALGO

Mathias Rav, Software developer

Mapping, understanding and mitigating flood risk in high resolution on a national scale

At SCALGO, we use cutting-edge massive terrain data-processing technology to create an interactive tool for mapping, understanding and mitigating flood risk. We have used decades of applied and theoretical research in computer science to build a technology that scales to country-wide setups on high-resolution (40cm) terrain data. Our tool answers a wide variety of flood risk queries on-the-fly and delivers detailed flood risk maps in the full resolution of data in a fraction of a second. At Local Rockstars, I will present the methods and technologies we use to produce a flood risk map based on DMI’s ocean level forecast, which is updated every six hours. Writing a national flood risk map in 40cm resolution takes two hours, but by preprocessing the terrain data to extract the important topological features, we are able to update the flood risk map in just a few minutes in this pilot study with DMI.

Mathias Rav speaker.png

 


10.50-11.20: Grundfos

Mikkel Haggren Brynildsen, Chief Data Scientist

Ontologies, Knowledge Graphs and Digital Twin Chatbots

Analytics & AI is a team of data scientists and developers in Grundfos. I work in the “lab” team of Grundfos A&AI where we do proof of concepts and pilots around industrial use of AI. In this talk, I will come across how our work in enabling stakeholders to interact with an IoT connected pump through a chat system using free text. The chat interface prototyped is a hybrid solution using external natural language processing (NLP) API’s (Microsoft LUIS), ontologies and rule based systems. In order to scale and to minimize role dependencies, we let the experts model knowledge and chat flow in a “web ontology language” (OWL) graph structure.

Mikkel grundfos speaker

 


11.25-11.55: Vestas Wind Systems A/S

Johnny Nielsen, Senior Specialist – Sensing & Information Fusion

SPEED UP SOFTWARE DEVELOPMENT WITH FLEET DATA AND MACHINE LEARNING

Developing and testing new software functions for wind turbines can be both difficult and time consuming. One of the main reasons is the need for the software to work on many different turbine types placed in very different climate conditions. Getting the software out for testing on a representative number of turbines can take several months and sometimes years. This talk will be about how to use Machine Learning with fleet data to develop functions that would otherwise be very difficult and very time-consuming to develop. Beside reducing development time, the approach can make sensors redundant while keeping or improving accuracy and reliability. Fleet data will also be used to validate the developed software functions on hundreds of turbines without installing it on a single one, reducing the verification time from months to days.

Johnny Nielsen

 


12.00-12.30: Kanda

Daniel Bang Rothmann, Machine Learning Engineer

Computer Vision at the Edge With Unity3D

Implementing solutions with machine learning is becoming more accessible, but getting models from development to the edge, running on a device, can be complicated! In this talk, I will demonstrate how we develop and run computer vision systems in industrial AR applications. The talk will cover a cloud-based deep learning workflow with PyTorch, considerations when deploying at the edge and tips to integrate, run and test models in Unity3D.

Daniel R speaker

 


13.55-14.25: Visiolink

Kaare Bøegh, CTO & Christian Pedersen, Senior Android Udvikler

Apps. Volume. Legacy. Headache.

Visiolink supplies the newspaper industry in Northern Europe with digital replica editions of the printed paper. We maintain 700+ apps for 200+ customers. The talk will give a very short overview of the many challenges associated with handling hundreds of apps. Some will be obvious, some less so. To illustrate how the sheer volume of apps creates complexities on its own, we will be taking a deep dive into the area of configurability, versioning and in particular how it is implemented on the Android platform.

Kaare speakerChristian speaker

 

 

 

 

 

 

 

 

 

 


14.45-15.15: Trifork

Stefan Veis Pennerup, ML Tech Lead

Classifying Toilets Using Deep Learning

Trifork excels at supporting clients in developing software solutions. For Pressalit, we have created a ML model that can identify toilet models and help their end users’ selection of a new toilet seat. In this talk, I’ll go through the fundamentals of a Convolutional Neural Network (CNN) and how we applied Transfer Learning to develop the final ML model, as well as deep dive into some of the challenges we faced when deploying it to production.

Stefan speaker

 


15.20-15.50: IT Minds

Mikkel Kaysen, Senior Software Developer

Ditching the virtual DOM

Doing state driven UI development on the web today, is almost always done using a framework that uses some kind of virtual DOM like React, Vue and Angular. But using a virtual DOM comes with a penalty of always re-rendering everything, unless explicitly told otherwise. In this talk, I will demonstrate some of the pitfalls one can easily fall into, when working with frameworks that uses virtual DOM, and how Svelte, by moving the grunt work from run time to build time, can solve these problems, and resolve in responsive UI without the complexity of declaring when to update.

Mikkel Kaysen It minds