ICTopen Intelligent Machines. Science meets business, March 22 2016
The program of the day will consist of a number of international invited speakers, and plenty time for networking during poster and exhibition sessions. The aim is to present a very comprehensive overview Dutch academic and industrial research.
Confirmed speakers
- Bert Kappen University of NijmegenOpening Track
11:30 - 12:10Max Welling University of Amsterdam
Deep machine learning
Deep learning has broken every record in terms of predictive accuracy over the last decade or so. What do we still need for this technology to become widely adopted in industry, healthcare, and government? In this talk I will briefly discuss the following aspects of deep learning for which we have made some progress recently: 1) visualization of a classification decision, 2) reduced energy consumption through spiking architectures, 3) reduced memory through automatic weight tying, 4) reliable uncertainty quantification through Bayesian modeling and 4) improved statistical strength through symmetry-group convolutions.
Prof. Dr. Max Welling is a research chair in Machine Learning at the University of Amsterdam and has secondary appointments as full professor at the University of California Irvine and as a senior fellow at the Canadian Institute for Advanced Research (CIFAR) . He is co - founder of “Scyfer BV” a university spin - off in deep learning. In the past he held postdoctoral positions at Caltech (’98 - ’00), UCL (’00 - ’01) and the U. Toronto (’01 - ’03). He received his PhD in ’98 under supervision of Prof. G. 't Hooft . Max Welling has served as associate editor in chief of IEEE TPAMI from 2011 - 2015 (impact factor 4.8). He serves on the board of the NIPS foundation since 2015 (the largest conference in machine learning) and has been program chair and general chair of NI PS in 2013 and 2014 respectively. He was also program chair of AISTATS in 2009 and will be a program chair of ECCV in 2016. He has served on the editorial boards of JMLR and JML and was an associate editor for Neurocomputing, JCGS and TPAMI. He received mu ltiple grants from Google, Facebook, Yahoo , NSF, NIH , NWO and ONR - MURI among w hich an NSF career grant in 2005 . He is recipient of the ECCV Koenderink Prize in 2010 and the best paper award at ICML 2012. Welling is currently the director of the master prog ram in artifi cial intelligence at the UvA, he is in the board of the Data Scien ce Research Center in Amsterdam and he co - directs the Qualcomm - UvA deep learning lab.
12:10 - 12:30Flash Presentations
14:00 - 14:40Thore Graepel Google Deep Mind, London
AlphaGo: Mastering the game of Go with deep neural networks and tree search
The game of Go has long been viewed as the most challenging of classic games for artificial intelligence owing to its enormous search space and the difficulty of evaluating board positions and moves. Here we introduce a new approach to computer Go that uses ‘value networks’ to evaluate board positions and ‘policy networks’ to select moves. These deep neural networks are trained by a novel combination of supervised learning from human expert games, and reinforcement learning from games of self-play. Without any lookahead search, the neural networks play Go at the level of state-of-the-art Monte Carlo tree search programs that simulate thousands of random games of self-play. We also introduce a new search algorithm that combines Monte Carlo simulation with value and policy networks. Using this search algorithm, our program AlphaGo achieved a 99.8% winning rate against other Go programs, and defeated the human European Go champion by 5 games to 0. This is the first time that a computer program has defeated a human professional player in the full-sized game of Go, a feat previously thought to be at least a decade away. The talk is based on a recent Nature article about work carried out at Google Deepmind in collaboration with David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, and Demis Hassabis.
Thore Graepel is an allround machine learning researcher with experience across statistical learning theory, kernel methods, optimization, Bayesian inference, graphical models, crowdsourcing and reinforcement learning. Before joining DeepMind, Thore spent twelve years at Microsoft Research developing machine learning algorithms at web scale including TrueSkill for player ranking in Xbox Live, AdPredictor for click-through rate prediction in Bing, and Matchbox for movie recommendations on Xbox Live marketplace. On the academic side, Thore is affiliated with the Computer Science Department of University College London as professor of machine learning. At DeepMind, Thore is returning to his original dream of helping to understand intelligence and build intelligent agents.
14:40 - 15:20Laurens van der Maaten Facebook, New York
Learning to Solve Vision without Annotating Millions of Images
The state-of-the-art performance on many computer-vision problems is currently held by systems trained on visual features that are produced by convolutional networks. These convolutional networks are trained on large supervised datasets that contain millions of manually annotated objects. Further improvements of these visual features will likely require even larger manually labeled data sets, which severely limits the pace at which progress can be made. In this talk, I will present work that explores the potential of leveraging massive, weakly-labeled image collections for learning good visual features. I present experiments in which we train convolutional networks on a dataset of 100 million Flickr photos and captions without any full supervision, and show that these networks produce features that perform well in a range of vision problems. Moreover, the experiments show that as a by-product of the learning, the networks appropriately capture word similarity and they can learn correspondences between words in different languages. This talk presents joint work with Armand Joulin, Allan Jabri, and Nicolas Vasilache.
Laurens van der Maaten is a Research Scientist at Facebook AI Research in New York. He is also affiliated to Delft University of Technology as an Assistant Professor. Previously, he worked as a post-doctoral researcher at University of California, San Diego. He received his PhD degree from Tilburg University. Laurens is on the editorial board of IEEE Transactions of Pattern Analysis and Machine Intelligence, and has served as an area chair for NIPS and ICML. In 2012, he received the SNN-STW Machine Learning Award for his work on dimensionality reduction. His research interests are in machine learning and computer vision.
This talk presents joint work with Armand Joulin, Allan Jabri, and Nicolas Vasilache.
16:00 - 16:40Christian Goerick Honda Research, Offenbach
Innovation Through Science
The Honda Research Institute Europe has the mission to innovate by combining scientific exploration and rigorous examination with the urge to create sustainable new technology. The talk will first present the scientific portfolio of HRI-EU and then discuss some concrete investigations in more depth. The domains covered will be automotive, motorcycle and robotics. The methods range from sensory processing and perception to planning and prediction.
Christian Goerick studied electrical engineering at the Ruhr-Universität Bochum, Bochum, Germany, and at the Purdue University, West Lafayette, IN, USA. He received a Dipl.Ing. degree in electrical engineering and a Ph.D. degree in electrical engineering and information processing from the Ruhr-Universität Bochum in 1993 and 1998. From 1993 to 2000 he was with the Institute for Neural Computation, Ruhr-Universität Bochum, where he became a Research Assistant, Doctoral Worker, and Project Leader in 1993, and a Post-Doctoral Worker, Project Leader, and Lecturer in 1998. The research was concerned with biologically motivated computer vision for autonomous systems and learning theory of neural networks. Since 2000 he is with the Honda Research Institute Europe, Offenbach, Germany, holding the position of a chief scientist. As a competence group leader he is responsible for system architecture and embodiment in the area of intelligent mobility.
16:40 - 17:20Onno Zoeter Xerox Research Grenoble (now at Booking.com, Amsterdam)
Machine Learning and Smart Cities: Parking Demand Management
On-street parking, just as any publicly owned utility, is used inefficiently if access is free or priced very far from market rates. Machine learning, dedicated sensors and control policies can be used to adjust rates to better match demand. The optimized rates encourage parkers to avoid peak hours and peak locations and reduce congestion and underuse. The solution we have developed is deliberately simple so that it is easy to understand, easily seen to be fair and leads to parking policies that are easy to remember and act upon. The algorithm is in use to change parking rates in over 6000 spaces in downtown Los Angeles since June 2012 as part of the LA Express Park project. Initial results are encouraging with a reduction of congestion and underuse, while more rates were decreased than increased. This work was done while the speaker was at Xerox Research.
Dr. Onno Zoeter is principal data scientist at Booking.com where he works with a very strong data science team on designing and improving prediction, filtering and ranking methods. Before joining Booking.com in 2015, Dr. Zoeter led the research on transportation demand management at Xerox’s European Research Lab and worked on large scale click-through rate prediction problems at Microsoft Research Cambridge. His research led to successful new products and received numerous awards. Among them the OECD International Transport Forum Innovation Award and the International Parking Institute Award of Excellence. Fortune selected him among its 20 Big Data All-Stars in 2014.
Based on our experience with earlier similar events (see www.ml2015.nl), we expect a large part of Dutch academic researchers to present their research at the symposium. In addition, the meeting tends to be well attended by industry. Companies are invited to present themselves with a stand. All academic researchers as well as researchers from industry and business are invited to present their latest application oriented research with a poster.
ICTopen Natural Artificial Intelligence, March 23 2016
The track Intelligent Systems is continued on 23 March with a scientific symposium that is concerned with the integration of two central topics: the scientific understanding of intelligence and the ability to replicate intelligence in engineered systems. Central questions of the symposium include how intelligence is grounded in computation, how these computations are implemented in neural systems, how intelligence can be described via unifying mathematical theories, and how we can build intelligent machines based on these principles. The program for the second day will present the initial results from the NWO program Natural and Artificial Intelligence, which funded 6 research projects in 2014. (see h http://www.nwo.nl/en/research-and-results/programmes/natural+artificial+ntelligence/projects):
11:30 - 11:50Tim de Bruin, Robert Babuska TUDelft
Deep Learning for Robust Robot Control
Recent years have seen a growing interest in the use of deep neural networks as function approximators in reinforcement learning. The Deep Learning for Robust Robot Control project aims to further develop these methods and apply them to robotic control applications. The first step in this direction has been to apply the Deep Deterministic Policy Gradient method for a robot control problem both in simulation and in a real setup. The importance of the size and composition of the experience replay database is investigated and some requirements on the distribution over the state-action space of the experiences in the database are identified. Of particular interest is the importance of negative experiences that are not close to an optimal policy. It is shown how training with samples that are insufficiently spread over the state-action space can cause the method to fail, and,how maintaining the distribution over the state-action space of the samples in the experience database can greatly benefit learning.
11:50 - 12:10Taco Cohen, Max Welling AMLAB, Informatics Institute, UvA
Symmetries and Equivariance in Deep Learning
Deep convolutional neural networks (CNNs, convnets) have proven to be very powerful models of sensory data such as images, video, and audio. They exploit translational symmetry in the data through convolutional operations. In this talk I will discuss how to extend convolutions to a larger group of symmetries, such as rotations, reflections etc. These "group-convolutions" result in representations that are "equivariant" to the larger group of symmetries and allow for a much higher degree of weight sharing resulting in improved statistical strength.
12:10 - 12:30Flash Presentations
14:00 - 14:20Moinuddin M. Haque, Paul Vogt, Afra Alishahi and Emiel Krahmer Tilburg University
Exploring patterns in children's interactions using neural network methods
Children while interacting, not only learn, but also adapt to their environment and communication partners. These interactions along with the linguistic information are richly augmented with social cues (such as eye gaze and gestures), which help to facilitate better interactions and learning. Based on the principles of cross situational learning (Quine 1960), computational models have learned to predict a word based on semantic features, using associative networks. The language game model of Steels (2003) is an example of an agent-based model in which agents interact with each other, exchange utterances and can learn from each other. Such models have been used to study language evolution. However, such models tend to implement interactions between agents using toy languages and thus do not reflect naturalistic interaction patterns. The CASA MILA (Vogt and Mastin 2013) corpus, which consists of longitudinal recordings of children interacting in naturalistic environments has been developed to overcome this problem (Vogt & Mastin 2013). The current paper presents a study to generate novel interactions based on observations from the corpus.
14:20 - 14:40Sander Bohte CWI Amsterdam
Networks of Spiking Neurons
With deep neural networks advancing AI in leaps and bounds, the actual computational paradigms of real neurons are being reexamined to possibly create novel types of artificial neural networks. The medium of neural communication in particular, by spikes, may help to advance critical elements of deep learning, such as (energy) efficient neural communication and asynchronous computation. Potential applications range from cell-phones capable of running large deep learning programs to efficient and graceful robot control. At the same time, networks of spiking neurons offer the most direct link to future brain computer interfaces. In this talk I outline some of the current trends and successes, as well as the current challenges faced in these endeavors and the relation to advances in neuroscience.
14:40 - 15:00Kimberley McGuire, Guido de Croon, Karl Tuyls, Bert Kappen
Exploration with a Swarm of Pocket Drones
A swarm of bees are quite magnificent if you look at them from a robotics point of view. They can detect & avoid obstacles, search for flowers, remember where they are and communicate that information to their neighbors. As a swarm they can explore a field of flowers within a short amount of time. To implement these capabilities into a swarm of pocket drones can be of benefit for many (indoor) surveillance applications. However, the current state of the art methods uses excessive computational power, sensing, memory and communication resources to enable low level control, high level navigation and swarm exploration. A micro aerial vehicle’s size is highly correlated with what it can transport and a pocket drone (which fits in the palm of your hand) cannot carry these resources. To use the state-of-the-art methods, it needs to be connected to an external computer to do the computations. The big challenge is have all the swarm exploration capabilities on-board of a pocket drone, dealing with strict hardware limitations. This project's goal is be to provide an efficient, nature-inspired solution that covers: (1) low-level navigation such as obstacle avoidance or flying through narrow corridors, (2) high level navigation allowing to reach places of interest, and (3) coordination with the other drones to explore the environment swiftly. In this presentation, I will focus most on the challenges involved in the first part, as I already obtained preliminary results allowing a pocket drone to stabilize itself with only on-board sensors. The first results are on detecting velocity and obstacles using computer vision with a low-weight stereo camera. Multiple variables can be detected with just one sensor, therefore makes other, single-purpose, sensors obsolete. The next step is for the pocket drone to efficiently manage its memory during exploration. Finally, multiple pocket drones will need to share minimal information within a certain range of each other, to make exploration decisions based on the actions of others. At the end of this project, a swarm of pocket drones can autonomously explore an indoor, GPS-deprived, building, as natural as a swarm of bees can.
15:00 - 15:20Dominik Thalmeier SNN
learning computations in spiking neural networks
Providing the neurobiological basis of information processing in higher animals, spiking neural networks must be able to learn a variety of complicated computations, including the generation of appropriate, possibly delayed reactions to inputs and the self-sustained generation of complex activity patterns, e.g. for locomotion. Many such computations require previous building of intrinsic world models. Here we show how spiking neural networks may solve these different tasks. Firstly, we derive constraints under which classes of spiking neural networks lend themselves to substrates of powerful general purpose computing. The networks contain dendritic or synaptic nonlinearities and have a constrained connectivity. We then combine such networks with learning rules for outputs or recurrent connections. We show that this allows to learn even difficult benchmark tasks such as the self-sustained generation of desired low-dimensional chaotic dynamics or memory-dependent computations. Furthermore, we show how spiking networks can build models of external world systems and use the acquired knowledge to control them.
16:05 - 16:30Award Ceremony & Closure
In addition, researchers are invited to present their work (see ICTopen call for posters. Some posters may be upgraded to oral presentation).Program Committee
The program of the Intelligent Systems track of ICTopen is organized by Bert Kappen (Radboud Universiteit), Paul Vogt (U van Tilburg), Tom Heskes (Radboud Universiteit), Robert Babuska (TU Delft), Guido de Croon (TU Delft), Sander Bohte (CWI), Pieter Roelfsema (NIN). The Intelligent systems track is part of the NWO annual ICTopen conference (22 and 23 March) which consists of 5 parallel tracks. The Intelligent systems track is coorganized by SNN, a non-profit organization that aims to promote research and applications on machine learning and artificial intelligence in the Netherlands.
Practical information
- Date and time: March 22, 2016 from 10.00 am until 6.00 pm.
- Venue: Theater de Flint, Amersfoort
- Registration (including lunch and reception): 100 Euro before February 16, 2016; 125 Euro later.
- Exhibition space including admission (including lunch and reception) for two persons: 1500 Euro (excl VAT). Exhibition space is 2.25 m2 stand space (1.5 x 1.5m). Includes poster wall, standup table and power outlet. Subject to availability, please inquire before February 27, 2016 at snn@science.ru.nl
- Deadline poster abstract submission: February 15, 2016. 1 poster per registration maximum.
Registration
Registration and poster submission through www.ictopen.nl
11:20 - 11:30