During the Midsummer week, the Swedish Society of Image Analysis (SSBA) hosted the first Swedish symposium in Deep Learningat KTH. The large amount of participants confirmed the great interest in machine learning among business and academics. From Savantic, Claes, Pontus, Miroslav and Karin participated and several companies were represented such as ContextVision, Tobii, Univrses and Apple.
Several interesting speakers were represented. First Cordelia Schmid et al. From INRIA, Grenoble talked about how to refine action detection, classification, and tracking in video (e.g. mobile phone speaker, drinking, leave bag at the airport, etc.) by integrating spatio-temporal information on classification and localization, instead of only training CNN per frame based on the video material. http://thoth.inrialpes.fr/research.php
Atsuto Maki summarized ongoing research at KTH, Perception and Learning Lab (RPL). In an exciting project, Atsuto looks with doctor Ali Ghadirzadeh, robotics professor Danica Kragic, etc. on how to use reinforcement learning to teach a robot with seven degrees of freedom to throw and receive balls from non-calibrated image data from a camera mounted on the robot.
Several speakers presented their work that are basically applications of known techniques, but some also worked on developing the technology itself:
Cristian Sminchisescu et al. From Lund develops a mathematical framework (will be integrated into TensorFlow) that allows integration of matrix functions such as SVD within a backpropagating deep learning network. Exciting with new "building blocks" for deep learning architectures.
Chetal Ningaraju from NVIDIA talked about how they contribute to the development of self-driving cars with specialized deep-learning hardware and architectures (Drivenet, Pilotnet). https://www.youtube.com/watch?v=HJ58dbd5g8g
The symposium ended with a panel discussion with interesting questions about the existence of Deep Learning and its development and clearly more freedom in the research, more money and better use-cases where on the agenda for better results. Savantic strongly recommends SSDL as a good gathering for Deep Learning in Sweden.