Enabling computers to understand images remains one of the hardest open problems in artificial intelligence. No machine vision system comes close to matching human ability at identifying the contents of images or visual scenes or at recognizing similarity between different scenes, even though such abilities pervade human cognition. In this talk I will describe research---currently in early stages---on bridging the gap between low-level perception and higher-level image understanding by integrating a cognitive model of perceptual organization and analogy-making with a neural model of the visual cortex.

While extracting fuzzy rules from data using exploratory analysis, there are many issues that must be addressed, particularly if we want to exploit the benefits of interpretability of fuzzy systems. There are many methods for generation of fuzzy rules, which may extract useful (in terms of accuracy) rules for high dimensional data also. But even for rules involving moderately large number of attributes, the main attraction of fuzzy systems cannot be exploited. These issues are akin to dimensionality reduction/structure identification and interpretability of the systems. We shall present an “interesting” mechanism to deal with some of these issues in an integrated manner. A unique attribute of this approach is that it can exploit subtle non-linear interactions between features, the problem (that we intend to solve), and the tool (that is used to solve the problem). The formulation is adapted to all three types of fuzzy systems: classification systems, Mamdani type systems and Takagi-Sugeno type systems. This approach can deal with necessary features, indifferent features, and derogatory features in an appropriate manner but may not minimize the use of redundant features. So, we further generalize the scheme so that it can control the redundancy in the selected features. The underlying philosophy can be easily adapted to other learning frameworks such as neural networks. The effectiveness of the approaches will be demonstrated using a set of applications.

Kernel methods have become one of the most widely used techniques in the field of machine learning. I will present my thoughts on what made them popular and where things are heading. I will discuss some recent developments for two-sample and independence testing as well as applications in different domains.

Mobile sensor systems and instrumented environments allow us to capture terabytes of context data on personal black boxes for people and products, which is enough to store complete digital lifelogs. Using data fusion, video and speech retrieval, machine learning, probabilistic reasoning, and plan recognition for information extraction from these digital shadows, semantically rich digital diaries become the basis of advanced web services for personalized assistance, healthcare, logistics, and safety. In this talk, we will show that sharing memories of smart products (ranging from cars to frozen pizzas) and their consumers in instrumented environments leads to innovative applications of the future internet of things and services, using our projects SharedLife and SemProm as examples. One of the key problems in generating digital diaries is managing the uncertainty in sensor, context and activity interpretation. We will present new methods for utility-based sensor selection and fusion and as well as state-aware activity and plan recognition that are based on a hierarchical combination of probabilistic automata, dynamic Bayesian networks and partially observable Markov decision processes and have been developed in my research group at DFKI for coping with uncertainty in lifelogging.

References:

Wahlster, W.; Feld, M.; Gebhard, P.; Heckmann, D.; Jung, R.; Kruppa, M.; Schmitz, M.; Spassova, L.; Wasinger, R. (2010): The Shopping Experience of Tomorrow: Human-Centered and Resource-Adaptive. In: Crocker, M. W.; Siekmann, J. (Eds.): Resource-Adaptive Cognitive Processes. Heidelberg: Springer, pp. 205-237.

Wahlster, W., Kröner, A., Schneider. M., Baus, J. (2007): Sharing Memories of Smart Products and Their Consumers in Instrumented Environments. In: it - Information Technology 50, 1, Oldenbourg, pp. 45-49.