Trust is good, but control is better: A speech by Assoc. Prof. Iasonas Kokkinos (CentraleSupelec / INRIA)
Deep Learning has been shown to yield excellent results across a range of Artificial Intelligence problems, including Computer Vision, Speech Recognition, and Natural Language Processing among others.
We will start this talk with a tutorial presentation of the ideas underlying Deep Convolutional Neural Networks (DCNNs), which provide a natural blend between signal processing and machine learning. We will illustrate some of the inner workings of these powerful classifiers and will present some of the visual tasks where they have been shown to be most successful.
Despite their huge successes, DCNNs are still largely considered as ‘black box' classifiers over which we may have limited control. We will therefore move on to presenting recent research efforts on integrating established computer vision ideas with DCNNs, thereby allowing us to incorporate task-specific domain knowledge in DCNNs.Through these works we will see how established techniques such as Multiple Instance Learning, Dense Conditional Random Fields, Gaussian Conditional Random Fields and Spectral Clustering can be integrated within deep architectures, yielding consistent improvements over generic, domain-agnostic baselines.
For more info have a look at the attached file
Contact: Professor P. Maragos (maragos@cs.ntua.gr)