#11: MLDublin meets ADAPT @ DogPatch Labs
- 1 minWe were thrilled to be back at DogPatch labs. Joined by: Deirdre Hogan, Majid Yazdani, Ahmed Selim, Chris Hokamp.
Painting style transfer is very popular nowadays. The idea is relatively simple: you provide a machine with a photograph and the machine paints it for you. While this works well with generic scenes such as landscapes, it is very challenging when dealing with selfies and head portraits. In this talk, Dr Ahmed Selim from the CONNECT Centre in Trinity College Dublin will present a machine learning technique for style transfer that specializes in head portraits. The technique uses deep learning models to combine information about the identity of an individual with information from the style of an artist to produce a faithful head portrait.
The goal of many Machine Learning tasks is to output a sequence of symbols. In many cases we actually know things about the output domain that could guide the search for optimal output, but this information can be difficult to incorporate at training time. In this talk, I'll discuss my recent work on structured search for models that output sequences, and propose a general method for constraining search when it makes sense to do so.