24. – 29. September 2017, Dagstuhl Seminar 17391
Deep Learning for Computer Vision
Auskunft zu diesem Dagstuhl Seminar erteilen
Simone Schilke zu administrativen Fragen
Andreas Dolzmann zu wissenschaftlichen Fragen
The paradigm that a machine can learn from examples much like humans learn from experience has fascinated researchers since the advent of computers. It has triggered numerous research developments and gave rise to the concept of artificial neural networks as a computational paradigm designed to mimic aspects of signal and information processing in the human brain.
There have been several key advances in this area including the concept of back- propagation learning (essentially gradient descent and chain rule differentiation on the network weight vectors) by Werbos in 1974, later popularized in the celebrated 1984 paper of Rumelhart, Hinton and Williams. Despite a certain success in pattern recognition challenges like handwritten digit classification, artificial neural networks dropped in popularity in the 1990s with alternative techniques such as support vector machines gaining attention.
With increasing computational power (and in particular highly parallel GPU architectures) and more sophisticated training strategies such as layer-by-layer pretraining, supervised backpropagation and dropout learning, neural networks regained popularity in the 2000s and the 2010s. With deeper network architectures and more training data, their performance has drastically improved. Over the last couple of years they have outperformed numerous existing algorithms on a variety of computer vision challenges such as object recognition, semantic segmentation and even stereo and optical flow estimation.
The aim of this Dagstuhl Seminar is to bring together leading experts from the area of machine learning and computer vision and discuss the state-of-the-art in deep learning for computer vision. During our seminar, we will address a variety of both experimental and theoretical questions such as:
- In which types of challenges do deep learning techniques work well?
- In which types of challenges do they fail? Are there variations of the network architectures that may enable us to tackle these challenges as well?
- Which type of network architectures exist (convolutional networks, recurrent networks, deep belief networks, long short term memory networks, deep Turing machines)? What advantages and drawbacks does each network architecture bring about?
- Which aspects are crucial for the practical performance of deep network approaches?
- Which theoretical guarantees can be derived for neural network learning?
- What properties assure the impressive practical performance despite respective cost functions being generally non-convex?
Creative Commons BY 3.0 DE
Daniel Cremers and Laura Leal-Taixé and Ian Reid and René Vidal
- Artificial Intelligence / Robotics
- Computer Graphics / Computer Vision
- Deep learning
- Convolutional networks
- Computer vision
- Machine learning