Magnus Axelsson

Data Scientist, Simula Consulting

Magnus is data scientist with a doctorate in Astrophysics. His professional experience is diverse, from data science at FINN.no, to machine learning consulting for Inmeta, to research in plasma chemistry as a theoretical plasma physicist at N2 Applied in the agri-tech industry. He currently works as a lead data scientist at Simula Consulting where he enjoys solving business problems using bleeding edge research techniques. When he’s not solving problems at work, he is most likely to be found wandering around the forest with his dog.

Sosiale medier

Alle foredrag med Magnus Axelsson

  • Program


  • 10:30  - 10:55
    Training neural networks (NNs) for production can be challenging, especially when data is limited. In these circumstances we want our model to reflect the fact that it cannot (or should not) be very confident about its predictions, as data tends to drift with time and thus lie outside the training data support. Sadly, research has shown that NNs are over-confident when applied to out-of-distribution data; this behavior is unintuitive and unwanted in applications where NN output is used as critical decision support.

    In this talk I will discuss how to regularize neural networks and make them more aware of the context they were trained in: I will discuss how to teach your NN to tell if incoming data is out of distribution, and at the same time yield better calibrated output. These and related methods will all be discussed in the context of a concrete business case that I recently worked on.
    #3 Smart




Share

Magnus Axelsson