SGAICO Annual Assembly and Workshop: Hot Topics in Cognitive Science & Artificial Intelligence

Loading Map....

Date(s) - 12/11/2015
09:00 - 16:30

EPFL Biotech Campus Geneva
Room: Batiment 1 - 6th Floor

The annual assembly and workshop of SGAICO will take place on November, 12 2015 on the Biotech Campus of EPFL in Geneva and is hosted by the BlueBrain Project.

The workshop will feature hot topics from Cognitive Science and Artificial Intelligence Research in Switzerland. Invited talks discuss recent and future developments in natural language understanding, heuristic search, and machine learning using recurrent neural networks. An interactive poster session provides additional insights into the Swiss research landscape covering a wide area of topics such as planning, robotics, transportation optimization and others.


Registration via until November 5, 2015.

Call for Participation

Our 2015 annual assembly will put Cognitive Science in the main focus. In the morning we will visit the Biotech Campus of EPFL at Geneva and attend a presentation about the Blue Brain project, followed by talks featuring Swiss research in AI and Cognitive Science. The afternoon is devoted to the discussion and exchange between participants and will include the SGAICO annual meeting as well as an interactive poster session featuring a wide range of AI-related topics.

On the evening before our assembly, we will meet for dinner in a restaurant in Geneva.

November 11, 2015: 19:00 Dinner at Restaurant Les 5 Portes, Rue de Zurich 8, 1201 Genève

November, 12 2015: Annual Assembly and Workshop 9 am – 4:30 pm Batiment 1, 6th floor

Here is a personal reflection on the event by Thilo Stadelmann.



8:30 – 9:00 Registration and Coffee


9:00 – 10:00 Welcome and Tour of the Biotech Campus – Marc-Oliver Gewaltig


10:15 – 11:00 Hervé Bourlard (IDIAP Martigny): ROCKIT: Roadmap for Conversational Interaction Technologies  

The EU ROCKIT ( Coordination and Support Action (CSA) aims at developing a research and technology transfer roadmap in multilingual and multimodal interaction, with a major emphasis towards increasing the innovation potential in industry, and SMEs in particular.

ROCKIT is concerned with the development of proactive, multimodal, social, and autonomous agents to enable rich, conversational interactions between people and between people and machines. Such agents should feature natural communication and interaction, proactive and autonomous behavior, social norms corresponding to the current environment, awareness of the current context, and a capability for social and affective behavior

Over the last 15 months, and based on more than 1,000 inputs collected from stakeholders from research and industry, ROCKIT has developed a first version of a research and innovation roadmap, mainly focusing on: (1) Societal drivers: mainly illustrated through 5 generic scenarios (adaptable interfaces for all, smart personal assistants, active access to unstructured information, communicative robots, and shared collaboration and creativity); (2) Trends of enabling technologies: what is available or to be developed, and when?; (3) Vertical markets: including tourism, health, manufacturing, security, collaboration support, etc.

The current version of the roadmap is accessible from and will be presented here, followed by discussions.


11:00 – 11:45 Renaud Richardet (EPFL): Large-scale extraction of brain connectivity from the neuroscientific literature

In neuroscience, as in many other scientific domains, the primary form of knowledge dissemination is through published articles in peer-reviewed journals. One challenge for modern neuroinformatics is to design methods to make the knowledge from the tremendous backlog of publications accessible for search, analysis and its integration into computational models.

In this talk, we introduce novel natural language processing (NLP) models and systems to mine the neuroscientific literature. We present an integrated NLP model to automatically extract brain region connectivity statements from very large corpora. This system relies on rule-based and machine-learning models (conditional random fields, support vector machines). It is applied to a large corpus of 25M PubMed abstracts and 600K full-text articles. Central to this system is the creation of a searchable database of brain region connectivity statements, allowing neuroscientists to gain an overview of all brain regions connected to a given region of interest. More importantly, the database enables researcher to provide feedback on connectivity results and links back to the original article sentence to provide the relevant context. The database is evaluated by neuroanatomists on real connectomics tasks (targets of Nucleus Accumbens) and results in significant effort reduction in comparison to previous manual methods (from 1 week to 2h). All models are freely available on Github.


11:45 – 13:00 Lunch Break


13:00 – 13:45 Malte Helmert (U Basel): Declarative Heuristics for State-Space Search

State-space search problems are ubiquitous in AI and other areas of computer science. Heuristic search with the A* family of algorithms is frequently used to address such problems, and much research in AI over the last decades has focused on developing better and better heuristics and furthering our theoretical understanding of existing heuristics.

This has led to increasingly more sophisticated approaches. A large amount of human ingenuity has gone into devising today’s advanced heuristics and into developing the underlying theory, especially in the case of optimal state-space search, where heuristics must satisfy a lower-bound property (be admissible) in order to be safe to use.

This talk explores the idea of shifting this burden of sophistication and ingenuity away from the human heuristic designer and onto the machines that perform the heuristic search. Equipped with appropriate reasoning algorithms and representations, computers can derive knowledge about state-space search more quickly, more reliably and in more complex settings than we humans are capable of. A heuristic search researcher or practitioner should be able to focus on which information a search algorithm ought to exploit, without having to worry about how to exploit it in the best way. Towards this end, the talk promotes a declarative approach to heuristic search that enables algorithmic reasoning about state spaces, solutions and heuristics in powerful and general ways.


13:45 – 14:30 Jan Koutnik (IDSIA Lugano): Introduction to Recurrent Neural Networks

We are currently experiencing a boom in machine learning, especially with neural network-based methods that go under the name Deep Learning (DL). Although the principals underlying much of DL were already in place a few decades ago, the availability, in the last 5 years, of affordable, massively parallel computing hardware in the form of graphical processing units (GPUs) has finally made these methods applicable to large class of real-world perceptual tasks. This talk will introduce recurrent neural networks (RNNs; the “deepest”networks) that are suitable for problems that involve processing of sequential data sets; primarily the Long Short-Term Memory network, which is currently the state of the art for speech recognition.


14:30 – 16:00 Interactive Poster Session & Coffee


Emeç Erçelik, EPFL: A CPG Based Locomotion on Robot Mouse

In this study, our aim is to reach a realistic mouse locomotion on the robot mouse which is modeled on Blender environment. At the beginning, we utilized a central pattern generator (CPG) as a controller for locomotion instead of a point neuron based spinal cord model for simplicity. In that way, we will be able to use the obtained locomotion to supervise the neuron based spinal cord model. The robot mouse consists of 12 servo based joints, the positions of which are controlled to provide locomotion. Each limb has three joints which are for thighs, knees and ankles. The CPG utilized is a modified version of Hopf oscillator. Each joint takes a position control value between 0 and 1 that is provided by CPG and has two parameters to scale and shift these values. The CPG also has 4 parameters which stand for oscillation convergence rate and frequencies of stance and swing phases of locomotion. The optimization of the parameters is performed with using Covariance Matrix Adaptation – Evolution Strategy (CMA-ES) evolutionary algorithm.


Thomas Koller and Tim vor der Brück, Hochschule Luzern: Shoe Track Search based on Hierarchical Features

Shoe tracks are important evidences on many crime scenes. On the one hand they can give important clues to the identity of the culprint, on the other hand they can be employed to draw a connection between several crimes. In order to establish if the same person is responsible for different crimes, the police tries to match the prints found in the different locations. However, to manually compare several hundreds or thousands of prints is a very tedious work and often not all images are available as the crimes occurred in different cantons or even countries. One way of facilitating the matching is to first find the corresponding shoe model by a search in a reference database. A fully automated system with short response times seems very difficult to come up with, as the quality of the images from the crime scenes is very bad. In order to facilitate searching the reference data base, we propose a new system where we assign hierarchical, attributed features describing specific characteristics to the shoe and reference tracks. The forensic expert is then able to input all of the data that he can perceive from the image, and the system will find the best matches. The new feature system explicitly allows to model uncertainties about the features by using a hierarchy. A fuzzy tree search will be developed to find the best matches quickly.


Jean-Daniel Dessimoz, HESSO.HEIG-VD: Cognition for Effective Control

For control, cognition is crucial in many respects. Six of the most significant aspects  will be discussed:
1. First, control implies the definition of a target state.
2. In some cases, a sequence of control actions, appropriately specified, may lead to the desired, goal state.
3. Actions are first defined in cognitive terms.
4. In general, systems are embedded in reality domains where unpredictable disturbances may occur, or worse, where much is yet unknown.
5. In closed-loop cases, some critical time properties of CS versus TBCS are required, to allow for success.
6. As control systems gain in scope, multiple agents appear and patterns of sociology must develop.



Marc Pouly, Roland Christen, Stefan Schnürle, Hochschule Luzern: Skyline Computation on Commercial Data

Many different skyline algorithms for preference based search have been proposed and compared in literature, but most of these evaluations were based on synthetic data. In this work, we present a case study of skyline computation on commercial data that we consider representative for many e-commerce platforms. The results of our measurements differ significantly from the results reported on synthetic data.


Wolfgang Schachner, CEPIAG: Emergence of spontaneous proto-cooperation

The emergence of proto-cooperation between our “baby” robot CEPIAGIII and its “mother” robot is the subject of our research, it focuses on the beginning of the construction of a social relationship. Its aim is to determine how it is possible to evolve from a proto-cooperation given by phylogeny to spontaneous proto-cooperation. This construction should gradually bring our “baby” robot to manage itself, that is to force it to schedule the right task at the right time. This could be the seeds of an architecture of the will.


Jana Koehler, Marc Tobler, Aburas Rustom, Hochschule Luzern: A Multi-Agent Architecture for Multi-Car Systems

The elevator industry has recently seen a renewed interest in so-called Multi-Car systems, which are systems of autonomous vehicles transporting passengers across a system of vertical and horizontal tunnels. This poster gives an overview on recent solutions proposed and presents an architectural case study where a system of intelligent agents implements a distributed and intelligent scheduling approach.


Silvan Sievers, Uni Basel: Factored Symmetries for Merge-and-Shrink Abstractions

Merge-and-shrink heuristics crucially rely on effective reduction techniques, such as bisimulation-based shrinking, to avoid the combinatorial explosion of abstractions. We propose the concept of factored symmetries for merge-and-shrink abstractions based on the established concept of symmetry reduction for state-space search. We investigate under which conditions factored symmetry reduction yields perfect heuristics and discuss the relationship to bisimulation. We also devise practical merging strategies based on this concept and experimentally validate their utility.


Thilo Stadelmann, ZHAW: PANOPTES – Automatic Segmentation of Newspaper Articles

We present results in automatically segmenting newspaper pages into disjunct articles. Image, text and OCR information will be fused by rule- and learning-based approaches to facilitate new products in media monitoring.


16:00 – 16:30 SGAICO Annual Meeting and Plans for 2016


16:30 Closing of the event