Abstract: It is well known that the performance of machine learning algorithms heavily rely on data representations. Deep learning algorithms attempt to learn multiple levels of representation with increasing abstraction. Such automatically learned data representations have been found useful in many fields including computer vision, robotics, and natural language processing. In this tutorial, I will walk you through the key ideas of deep learning and cover the basics of deep neural networks. I will also briefly introduce some widely used deep learning models such as Deep Belief Networks and auto-encoders, together with their applications in computer vision and robotics.
Abstract: A tour of Caffe with practical tips on issues such as setting up your data, choosing architectures and hyperparameters, modifying caffe, selecting the best GPU hardware etc.
Abstract: Convolutional neural networks have achieved considerable success in many tasks in computer vision such as image classification, object detection / recognition or semantic segmentation. These networks are computationally demanding and not always feasible for embedded platforms where power and computational resources are relevant. Recent works have shown significant redundancy in the parameters of these networks. This over parametrization seems necessary to overcome the challenges existing in highly non-convex optimization problems. In this talk I review recent techniques to speed up and reduce the parameter redundancy existing in current networks.
Cesar Cadena: Depth Estimation with Multi-modal Auto-Encoders
Daniel Weimar: Context-aware Deep Convolutional Neural Networks for Industrial Inspection [Slides]
Abstract: Autonomous robots are faced with a series of learning problems to optimize their behavior. In this presentation I will describe recent approaches developed in my group based on deep learning architectures for object recognition and body part segmentation from RGB(-D) images and terrain classification from sound. In addition, I will present an approach using sparse coding to compactly represent three-dimensional environments. For all approaches I will describe expensive experiments quantifying in which way the corresponding algorithm extends the state of the art.
Abstract: Structured output learning concerns the problem of predicting multiple variables that have dependency, with Conditional random field (CRF) as a typical example. It shows great promise in tasks like semantic image segmentation. Recently, there is mounting evidence that features from deep convolutional neural networks (CNN) set new records for various vision applications. Here I show how we can combine CRFs with deep CNNs to predict complex labels while considering the dependencies between the output variables. The first application is to learn depth from single monocular images. Compared with depth estimation using multiple images such as stereo depth perception, depth from monocular images is much more challenging. We propose a deep structured learning scheme which learns the unary and pairwise potentials of continuous CRF in a unified deep CNN framework, termed Deep Convolutional Neural Fields. For the second application, we proffer a new, efficient deep structured model learning scheme, in which we show how deep Convolutional Neural Networks (CNNs) can be used to estimate the messages in message passing inference for structured prediction with CRFs. With such CNN message estimators, we obviate the need to learn or evaluate potential functions for message calculation. This confers significant efficiency for learning, since otherwise when performing structured learning for a CRF with CNN potentials it is necessary to undertake expensive inference for every stochastic gradient iteration. We also demonstrate that it yields results that are competitive with the state-of-the-art in semantic segmentation for the PASCAL VOC 2012 dataset.
We will have the following poster presenters: Edison Guo, Fahimeh Rezazadegan, Fangyi Zhang, Frederic Maire, James Sergeant, Peter Anderson, Rodrigo Santa Cruz, Sean McMahon, Daniel Weimar
We want to congratulate Cesar Cadena, who won the best paper award, voted by the attendees and the organisers!
Cesar, enjoy your Titan X!!