UCA Deep Learning School 2018

Program

Program Deep Learning @UCA Event 2018, From June 25 to 29
• 3-hour Lecture in the morning (9 am – 12:30)
• Coffee break (around 10:30)
• Poster session at lunch break (1 pm to 2pm)
• Optional: Nvidia Lab sessions (2pm to 5pm)
• Coffee break (around 3:30pm)

(Warning! From this year, DL@UCA requires registration fees in some cases. The details can be found below)

CSS

June 25th

Deep Learning for Signal Processing in Health and Wellbeing by Professor Björn Schuller

Deep learning is increasingly substituting traditional signal processing methods given the recent advances in end-to-end learning even from raw signals and often impressive improvements in performance. In this lecture, we will be going along the chain of processing in a typical pattern recognition system from signal capture to decision making discussing opportunities for deep learning algorithms to be entered. This will include the pre-processing and denoising of signals such as by auto encoders. Then, feature representation learning by convolutional layers will be featured. The optimal modelling of features will include recurrent memory-enhanced topologies such as long-short-term memory and gated recurrent units. To best cope with sparsity of (labelled) learning data, active and semi-supervised learning with human annotators in the loop, transfer learning across domains, and generative adversarial topologies will form a further major point of interest. Then, we will deal with how to best self-optimize such systems in the sense of “automatic machine learning”. Further aspects of discussion will include efficiency such as energy-awareness for mobile application, and interpretation of learnt models. The methods presented will be applicable to a large variety of intelligent signal processing tasks. However, to exemplify matters in a coherent way, application use-cases will be focusing on the health and wellbeing domain.

CONFERENCE:
Deep Learning for Signal Processing in Health and Wellbeing by Pr Björn Schuller
LABS:
Image Classification with DIGITS – as it is the basic for most of the other LABS
Signal Processing using DIGITS – works with DIGITS CNN

June 26th

Budgeted learning and Adverse model disentanglement by Professor Ludovic Denoyer

Despite the success of deep learning techniques for predictive problems only if provided with large amounts of labeled data, we are still far from being able to the create intelligent agents.
Particularly, if we want them to interact with people, computers cannot handle human constraints like communication through a common language, low-frequency interactions, use of common and meaningful elements, reasoning, etc...
The presentation will focus on two topics: the first part of the talk will be on budgeted learning models where the objective is to develop systems that automatically takes into account operational constraints (CPU consumption, Memory condumption, etc...). The second part will be focused on the problem of disentanglement where the objective is to identify underlying latent factors responsible of the observed data.

CONFERENCE:
Budgeted learning and Adverse model disentanglement by Pr Ludovic Denoyer
LABS:
Neural Network Deployment with DIGITS and TensorRT – as it is about the practical implication to put an AI to live (Budget, Resources)
Image Captioning by combining CNN and RNN – first round for RNN. Saw on the event web site that there is no track related to RNN

June 27th

Deep Learning for Autonomous Vehicles by Professor Li Erran Li

Despite recent advances, major problems in autonomous driving are far from solved—both in terms of fundamental research and engineering challenges. Machine learning holds the key to solving these challenges. I will explore machine learning fundamentals for autonomous driving and discusses current work in progress. Specifically, I will cover machine learning for perception (3D object detection, tracking, semantic scene understanding), prediction, planning and control (imitation learning, deep reinforcement learning for driving policies).

CONFERENCE:
Deep Learning for Autonomous Vehicles by Pr Li Erran Li
LABS:
Object Dection with DIGITS – Object Detection as the base for autonomous
Image Segmentation with Tensorflow – the other base discipline for autonomous

June 28th

Deep Reinforcement Learning by Professor Li Erran Li

Deep reinforcement learning has enabled artificial agents to achieve human-level performance across many challenging domains (for example, playing Atari games and Go). I will cover several important algorithms, including deep Q-networks and asynchronous actor-critic algorithms (A3C), discusses major challenges, and explores promising results for making deep reinforcement learning applicable to real-world problems in robotics and natural language processing.

CONFERENCE:
Deep Reinforcement Learning by Pr Li Erran Li
LABS:
Word Generation with Tensorflow – As we have for Reinforcement Learning, we go forward with RNN

June 29th

Training object localization models in weakly supervised settings by Professor Vittorio Ferrari

A key goal of computer vision is to interpret complex visual scenes, by recognizing visual concepts, localizing them, and understanding their interactions within the scene. To achieve this we need powerful visual learning techniques to acquire rich models capturing the diversity of the visual world. One crucial ingredient is the ability to learn visual localization models with minimal human supervision. This is necessary to scale to a large number of concepts and many training samples. In this lecture I will give an overview of weakly supervised techniques, that can learn without any location annotation, given only image-level labels. I will also talk about recent techniques in an intermediate regime where humans provide partial location information through answering simple questions, such as clicking in the middle of an object or verifying annotations produced by the learner. These seem to hit a particularly sweet spot in the trade-off between quality of the learned models and human supervision time.

CONFERENCE:
Training object localisation models in weakly supervised settings by Pr Vittorio Ferrari

Registration

Access & Accommodation

Accommodation facilities for academic members:
The University provides student individual rooms and studios in Nice at very interesting prices.
Please visit the following site
Résidences ICARE - Crous de Nice-Toulon
For English version, go to Bottom of the page and click on Langage.

Special sessions

Poster sessions will be held during lunch time.
Submitted posters will be selected by DL@UCA scientific committee.
One free entrance to the DL event will be granted for one of the authors of each accepted poster.

++resource++unice.templates.stylesheets/onepage/hero75.css

++resource++unice.templates.stylesheets/onepage/hero_image_only.css