Lecture | Lunch included
CANCELLED: Florence Nightingale Colloquium presents Max Welling
- Friday 20 March 2020
- The seminar is targeted at a broad audience, in particular we invite master students, PhD candidates and supervisors interested or involved in the Data Science Research programme as well as colleagues from LIACS and MI to attend. The seminar is organized by the DSO, MI and LIACS.
- Florence Nightingale Colloquium
Niels Bohrweg 1
2333 CA Leiden
- 407 - 409
Link to registration form.
Deep learning has been amazingly successful in solving AI tasks where the domain is well defined and a lot of data is available for that domain. Examples include speech recognition, (medical) image analysis, automatic translation and so on. However, increasingly researchers are worried about 'out-of-domain generalization' where a model needs to perform well on data that the learning algorithm has never seen before. Traditional engineering approaches, which include a large amount of prior knowledge and model the physical data generation process often do better in this regime. Therefore, a natural question to ask is how to combine these two approaches. In this talk we introduce 'neural augmentation', which builds a black box neural network around an interpretable graphical (generative) model of the data. The model is trained end-to-end by unrolling the inference procedure in the graphical model (given by belief propagation), augmenting it with a factor-graph neural network, and backpropagating through the whole recurrent architecture so defined. We illustrate these ideas on LDPC error correction decoding with bursty channels and on reconstructing MRI images from a reduced set of (fourier) measurements.