Abstract

Robotics is breaking free from constrained environments to physically interact with a variety of objects and humans. This advanced level of interaction can be supported by multi-sensor systems capable of perceiving the objects around them, changes to the environment, and the actions of human users. Machine touch, in particular, can play a key role within the context of a multi-sensor object recognition system, enabling robustness to illumination changes and occlusions, extraction of object properties, etc. The majority of tactile recognition systems are based on grasping, however, recently, object contact through a continuous tactile sensor, i.e., artificial skin, has emerged as a powerful alternative.

In this lecture, we will discuss artificial skin sensing using Electrical Impedance Tomography (EIT). For a set of typical EIT measurements, one or more rings of electrodes are placed around the medium that an object is positioned. A small current is injected into an electrode pair and the voltages between the other pairs of electrodes are recorded. EIT is capable of simultaneously detecting multiple objects, while providing a richer, i.e., image-based, description of their properties. This offers a number of advantages, e.g., low cost, wire-free, good temporal resolution, and, in the context of tactile sensing, a straightforward interpretation of pressure information. Specifically, when pressure is applied to a pressure-sensitive artificial skin, the surface of the skin deforms, resulting to changes in its conductivity distribution, which can be represented as an image through EIT.

To empower EIT for tactile sensing, machine learning is used to solve the complex problem of image reconstruction through patch-based decomposition. Determining the location of the tactile stimulus is posed as an image segmentation process, which is addressed using transfer learning. Finally, features extracted from the segmented objects are used for classification with state-of-the-art machine learning methods.

Short Biography

Dr. Panos Liatsis is a Professor in the Department of Electrical Engineering and Computer Science at Khalifa University of Science and Technology. Prior to joining Khalifa University, he was Professor of Image Processing and Head of Department of Electrical and Electronic Engineering at City, University of London.

He received the Diploma degree in Electrical Engineering from the University of Thrace in Greece and the Ph.D. degree in Electrical Engineering & Electronics from the University of Manchester, UK.

Dr. Liatsis' research interests include machine learning, pattern recognition, and computer vision, with applications in biomedical image and signal processing, security, and intelligent transportation systems.

He published over 200 research contributions in high-impact factor journals, books and international conference proceedings. He is on the editorial boards of international journals and a member of the International Program Committees of international conferences in the fields of image processing, machine learning and intelligent systems.

Abstract

Over the last half decade, the computer vision field has made huge strides on core image understanding tasks such as object recognition, detection, segmentation and keypoint estimation -- today these are technologies that are ubiquitously available to everyday consumers. However, video understanding technology is relatively less pervasive as computer vision models must deal with more difficult appearance, less labeled data, and computational challenges. In this talk, I will cover some of what I consider to be the core challenges of computer vision for video understanding, discussing these in the context of several recent research projects by the Perception team at Google.

Short Biography

Jonathan Huang is a senior research scientist at Google and currently works on deep learning for machine perception. He received his M.Sc degree and Ph.D from the School of Computer Science at Carnegie Mellon University in 2008 and 2011 respectively. From 2011 to 2014 he was an NSF Computing Innovation (CI) postdoctoral fellow at the geometric computing group at Stanford University where he also received his B.S. degree in Mathematics in 2005. His research interests lie primarily in deep learning, and probabilistic reasoning with combinatorially structured data with applications in computer vision and online education. To see a list of publications and projects, visit http://www.jonathan-huang.orga>.

Abstract

Visual scene understanding involves estimating quantities of interest (such as object category, location, etc.) from observed images and videos. Bayesian approaches to such problems are gaining importance as these provide uncertainty estimates of the variables being estimated. However, these tend to be expensive and unfeasible for real-time implementation. We present a framework for efficiently solving such problems in real-time using probabilistic analysis-by-synthesis (Approximate Bayesian Computations). The underlying generative models needed for ABC are built from realistic off-the-shelf simulation software but additionally include a simple Bayesian error model for the gap between simulation outputs and real data. We achieve real-time performance by introducing two novel approaches: neural emulators and tree pyramidal importance sampling.

Short Biography

Dr. Javier Felip Leon is a Research Scientist at Intel Labs since 2016. Javier completed his Ph.D. on Computer Science at Universitat Jaume I (Spain) in early 2016, where his research focused on Robotic manipulation under uncertainty (some robot videos here). Currently, his research focuses on analysis-by-synthesis, applied probabilistic methods, sampling methods and intuitive physics. Javier has authored more than 20 peer-reviewed paper publications in conferences and journals such as ICRA, IROS, HUMANOIDS and RAS.

Abstract

In recent decades, advances in capture devices and the increase of available digital image data have stimulated the creation of methodologies for data processing that produce various forms of valuable models, such as descriptors, classifiers, approximations and visualizations. These models are often developed in the field of machine learning, which is characterized by a large number of available algorithms. These algorithms often do not have guidelines to identify the most appropriate one based on specific data to which they will be applied and the nature of the problem under analysis.

There is a knowledge that allows to relate the features of the algorithms and data that present a good performance to fulfill a specific task, known as Meta-Knowledge, which can include information on algorithms, evaluation metrics to calculate similarity of datasets or relation of tasks. Being Meta-Learning the study of methods based on principles that explore the Meta-Knowledge to obtain efficient models and solutions, adapting the processes of Machine Learning and Data Mining.

The research carried out in this work analyzes the applications and advantages offered by Meta-Learning in the field of digital image processing. To carry out this task, different types of images, characterizers, and feature analysis techniques are used.

The results obtained show that methodology based on Meta-Learning is efficient when applied in the processing of digital images for identification and storage of experience generated by developing methodologies for classification of different types of images, obtaining a high performance with respect to an evaluation metrics. This statement means that Meta-Learning allows recommending the most appropriate methodology to perform the processing of a specific type of image based on features of the dataset under analysis and the type of specific task to be performed.

Short Biography

Aristófanes Corrêa Silva received a Ph.D. degree in Informatics from Pontifical Catholic University of Rio de Janeiro – Brazil in 2004. Currently, he is a Full Professor at the Federal University of Maranhão (UFMA), Brazil. He teaches image processing, pattern recognition and programming language. His research interests include image processing, image understanding, medical image processing, machine vision, artificial intelligence, pattern recognition and, machine learning. Prof. Aristófanes has authored more than 80 peer-reviewed paper publications in scientific journals.

Short Biography

Luis Fernando Marin was a Ph.D. student in computer science at the Pontifical Catholic University of Rio de Janeiro (PUC-RIO). He received his engineering degree from the University of Antioquia in Colombia and his master's degree from the Federal University of Maranhão, both with an emphasis in computer science. His research interest includes image processing, image understanding, medical image processing, machine vision, artificial intelligence, pattern recognition and, machine learning.

Abstract

Machine learning can be regarded as computational techniques for learning probability distributions from data. Interpretable machine learning refers, however, to methodologies that make the learned information understandable to humans. In this tutorial, It will be overviewed the interpretability concepts and methods and their application to lawsuit prediction in the energy sector.

Short Biography

André B. Cavalcante is professor at Federal University of Maranhão, Brazil. He received his Doctorate degree in Information Science in 2015 from Nagoya University, Japan. His research interest includes unsupervised generative learning, and its applications to vision.