In this position paper we argue that the concepts of memory and computation as they are commonly used in theories of cognition are strongly influenced by our intuitive understanding of the corresponding concepts in contemporary computer systems, leading to an implicit loss of biological plausibility. To support our argument we provide an alternative perspective on memory and computation that allows a closer comparison of the capabilities of computer programs running on computer systems and neurobiological systems showing that computer programs exhibit a computational flexibility that is difficult to reconcile with neurobiological constraints. We end this paper by offering a neurobiologically plausible perspective on memory that views memory as a dynamic, distributed process that is an intrinsic part of a neurobiological network that integrates information, e.g., sensory information.

In robot perception, as well as in other areas of 3-D computer vision, keypoint detection is the first major step for an efficient and accurate 3-D perception of the environment. Thus, a fast and robust algorithm for an automatic identification of keypoints in unstructured 3-D point clouds is essential. The presented algorithm is designed to be highly parallelizable and can be implemented on modern GPUs for fast execution. The computation is based on a convolution of a voxel based representation of the point cloud and a voxelized integral volume. The generation of the voxel-based representation neither requires additional surface information or normals nor needs to approximate them. The proposed approach is robust against noise up to the mean distance between the 3-D points. In addition, the algorithm provides moderate scale invariance, i. e., it can approximate keypoints for lower resolution versions of the input point cloud. This is particularly useful, if keypoints are supposed to be used with any local 3-D point cloud descriptor to recognize or classify point clouds at different scales. We evaluate our approach in a direct comparison with state-of-the-art keypoint detection algorithms in terms of repeatability and computation time.

We propose an adaptive approach to 3-D object classification. In this approach appropriate 3-D feature descriptor algorithms for 3-D point clouds are selected via reinforcement learning depending on properties of the objects to be classified. This approach is supposed to be able to learn strategies for an advantageous selection of 3-D point cloud descriptor algorithms in an autonomous and adaptive way, and thus is supposed to yield higher object classification rates in unfamiliar environments than any of the single algorithms alone. In addition, we expect our approach to be able to adapt to subsequently added 3-D feature descriptor algorithms as well as to autonomously learn new object categories when encountered in the environment without further user assistance. We describe the 3-D object classification pipeline based on local 3-D features and its integration into the reinforcement learning environment.

A widely adopted approach to study and understand complex systems such as ant colonies or economic systems consists in their modeling and simulation. In contrast to the predominat use of models and simulation in science as a substitute for the real system in order to predict the system’s beaviour, the methodology of exploratory modeling uses modeling and simulation as a tool to increase knowledge and understanding of the systems themselves, for example to better understand the dynamic properties of a system. In this paper we propose a model which was specifically devised to support this process of exploratory modeling. The model defines four lightweight building blocks such as information processing entities that can be freely combined to model a particular complex system. Furthermore, the model provides an explicit state representation that comprises the entire model including an explicit representation of the information that is individually available to every information processing entity of the model. We illustrate our introduction of the proposed model by means of short examples of a concrete system for the simulation of motions in a flock of birds.

In this paper we present a new computational model for grid cells. These cells are neurons in the entorhinal cortex of the hippocampal region that encode allocentric spatial information. They possess a peculiar, triangular firing pattern that spans the entire environment with a virtual lattice. We show that such a firing pattern can emerge from a dendritic, self-organized learning process. A key aspect of the proposed model is the hypothesis that the dendritic tree of a grid cell can behave like a sparse self organizing map that tries to cover its input space as best as possible. We argue, that the encoding scheme used by grid cells is possibly not limited to the description of spatial information and may represent a general principle on how complex information is encoded in higher level brain areas like the hippocampal region.

One aspect of self-organizing systems is their desired ability to be self-learning, i.e., to be able to adapt dynamically to conditions in their environment. This quality is awkward especially if it comes to applications in security or safety-sensitive areas. Here a step towards more trustful systems could be taken by providing transparency of the processes of a system. An important means of giving feedback to an operator is the visualization of the internal processes of a system. In this position paper we address the problem of visualizing dynamic processes especially in self-learning systems. We take an existing self-learning system from the field of computer vision as an example from which we derive questions of general interest such as possible options to visualize the flow of information in a dynamic learning system or the visualization of symbolic data. As a side effect the visualization of learning processes may provide a better understanding of underlying principles of learning in general, i.e, also in biological systems. That may also facilitate improved designs of future self-learning systems.

Computer vision can highly benefit from modern learning methods. In the context of an active vision environment we introduce a machine learning approach which is able to learn strategies of object acquisition. We propose a hybrid learning method, called Sphinx, that combines two approaches originating from seperate disciplines of computer science, namely reinforcement learning on the one hand and belief revision on the other. The former represents knowledge in a numerical way, while the latter is based on symbolic logic and allows reasoning. Sphinx is designed according to human cognition and interacts with its environment by rotating objects depending on past perceptions to acquire those views which are advantageous for recognition. Our method was successfully applied in simulations of object categorization tasks. 1

From a conceptual point of view, belief revision and learning are quite similar. Both methods change the belief state of an intelligent agent by processing incoming information. However, for learning, the focus in on the exploitation of data to extract and assimilate useful knowledge, whereas belief revision is more concerned with the adaption of prior beliefs to new information for the purpose of reasoning. In this paper, we propose a hybrid learning method called SPHINX that combines low-level, non-cognitive reinforcement learning with high-level epistemic belief revision, similar to human learning. The former represents knowledge in a sub-symbolic, numerical way, while the latter is based on symbolic, non-monotonic logics and allows reasoning. Beyond the theoretical appeal of linking methods of very different disciplines of artificial intelligence, we will illustrate the usefulness of our approach by employing SPHINX in the area of computer vision for object recognition tasks. The SPHINX agent interacts with its environment by rotating objects depending on past experiences and newly acquired generic knowledge to choose those views which are most advantageous for recognition.

Images play an important role in visualization. As users are more willing to adopt a product if it evokes pleasurable feelings the aesthetic appeal of interfaces becomes more important. Thus, there is a growing need to generate also images which appear aesthetically to the user. Starting with the modularities of the human visual system, we derive six dimensions of visual aesthetics. For each dimension we explore, inspired by principles of the visual arts and insights of cognitive neuroscience, which pecularities of the dimensions are particularly adequate for an aesthetic impression. Accompanied by a fair number of image examples, these considerations result in an easy to understand guideline for computer scientists and interface designers how to deal with images in terms of aesthetics.

We propose an architectural model for a responsive vision system based on techniques of reinforcement learning. It is capable of acquiring object representations based on the intended application. The system can be interpreted as an intelligent scanner that interacts with its environment in a perception-action cycle, choosing the camera parameters for the next view of an object depending on the information it has perceived so far. The main contribution of this paper consists in the presentation of this general architecture which can be used for a variety of applications in computer vision and computer graphics. In addition, the funcionality of the system is demonstrated with the example of learning a sparse, view-based object representation that allows for the reconstruction of non-acquired views. First results suggest the usability of the proposed system.

Pictures taken by a rotating camera cover the viewing sphere surrounding the center of rotation. Having a set of images registered and blended on the sphere what is left to be done, in order to obtain a flat panorama, is projecting the spherical image onto a picture plane. This step is unfortunately not obvious - the surface of the sphere may not be flattened onto a page without some form of distortion. The objective of this paper is discussing the difficulties and opportunities that are connected to the projection from viewing sphere to image plane. We first explore a number of alternatives to the commonly used linear perspective projection. These are 'global' projections and do not depend on image content. We then show that multiple projections may coexist successfully in the same mosaic: these projections are chosen locally and depend on what is present in the pictures. We show that such multi-view projections can produce more compelling results than the global projections

Ullman and Basri [1] have shown theoretically, that a three-dimensional object can be represented by a linear combination of two-dimensional images of the object. But they have applied their calculations to artificially created images only, like line drawings of cars. The application to images of real objects turns out to be difficult, because a crucial point of their algorithm is the knowlegde of correspondences in the sample views. In this article we describe a biologically inspired system which automatically provides correspondences between views of a three-dimensional object. This enables us to apply Ullman and Basri's linear combination approach to images of arbitrary, real objects. We give detailed formula of our linear combinations and examples for reconstructed object views.

The viewing hemisphere of a 3-dimensional object can be partitioned into areas of similar views, termed view bubbles. We compare two procedures of generating view bubbles: tracking of object features, i.e., Gabor wavelet responses, by utilizing the continuity of successive views and matching of features in different views, which are assumed to be independent. Both pro-cedures proved to be appropriate to detect canonical views. We found no difference concerning the size of the view bubbles, but tracking provides more precise correspondences than matching. Tracking is more appropri-ate for recognizing changes of features, whereas matching is more suitable if features of the same appearance are to be found. 1 Subject of Investigation For pose-invariant object recognition and pose estimation of objects it is necessary to uti-lize an appropriate object representation. A naive representation might consist of densly spaced views of an object’s viewing sphere. Our aim is to reduce such a “full ” represen-