Docentship lecture: Rethinking data models in learning

The Department of Electrical Engineering hereby invites all interested to a docentship lecture in subject Electrical Engineering with Specialization in Signal Processing.

Lecturer: Dr. Ayca Ozcelikkale

Title: Rethinking data models in learning

Date: Wednesday 8 September 2021 at 13.00


Chairperson: Professor Anders Ahlén

Representative of the Docentship Committee: Professor Jan Isberg


Machine learning (ML) techniques have shown extraordinary success in a wide range of application areas, such as autonomous driving, face recognition and language translation.   An important aspect in the modern ML models is possible overparametrization, i.e. the number of parameters in the model is larger than the number of the samples in the training data.  It is traditionally stated that overparametrization and near-perfect fit in training leads to poor performance on new data, and hence poor generalization performance. Strikingly, modern ML methods can achieve state-of-art performance even in the highly overparametrized regime.   This seemingly surprising behavior has recently been described with double descent curves highlighting that generalization error can decrease with increasing number of model parameters even in the overparametrized regime. In this lecture, we will discuss the classical perspective on overparametrization as well as the double descent phenomena using linear models.  We will present applications in feature selection as well as distributed learning. We will reveal several potential pitfalls for data collection and learning; and provide guidelines for good practice.  

The lecture is an obligatory teaching test for those applying for admittance as docent and it should be possible for students and others with basic academic education in the relevant field to follow it. The lecture lasts 40-45 minutes with subsequent discussion. The lecture will be given in English.