Challenges in real-world applications of neural networks
Sprache des Titels:
Neural network based systems are now state-of-the-art for a large
variety of tasks, often being orders of magnitude better than competing
methods. However, many such results are obtained on large, well-cured standard data corpora, and may not always generalize well to novel types of data, or to novel tasks. In particular, tasks which relate on biosignals (i.e. physiological signals collected from the human body) come with their own set of challenges and constraints, including small corpus size, large data variability, recording artifacts, and missing annotations. In this discussion-style talk, I will present own work on some of these issues and outline possible remedies, as well as open questions and necessary steps to tackle them.