Sciences Informatiques : IMRA à l'UCA Deep Learning School 2019
From July 15th to July 19th, the Université Côte d’Azur organized its 3rd Summer School in Deep Learning with some of the best academic actors in the Deep Learning field. One member of the Computer Science team of IMRA attended the morning lectures.
(Image Copyright Université Côte d'Azur)
Of particular interest for IMRA CS team's activities were the lectures from
- Prof. Martial Hebert (lecture title: “Research Challenges in Using Computer Vision in Robotics Systems”) who presented very challenging topics such as Introspection (the ability of a system to detect when it is failing), Low-Shot Learning (the ability of a system to learn from very few annotated samples) and Self-Supervision (the ability of a system to learn without any annotated samples).
- Prof. Alexandre Alahi (lecture title: “Deep Learning for Self Driving Cars and Beyond”) who highlighted the importance of (future) Prediction as one of the 3 fundamental pillars of self-driving car technology (along with Perception and Planning). He presented his work on forecasting human trajectories with social interactions (‘social LSTM’ and ‘Social GAN’) and cited multimodal (i.e. multiple future) prediction as a very important research topic.
- Prof. Graham Taylor (lecture title: “Towards Interpretable and Robust Machine Learning Systems”) who talked about the influence of Batch Normalization on the robustness to adversarial attacks (increased accuracy leads to reduced robustness) and introduced issues related to calibration, confidence and uncertainty estimates (e.g. how the ‘probabilities’ usually output by NN classifiers relate to their true correctness likelihood, how to perform Ouf Of Distribution Detection, how to improve Reinforcement Learning with the uncertainty).
For more information, http://univ-cotedazur.fr/events/deep-learning-school/2019