KiMeKo (KI-Med-Kollaborationsplattform) is a publically funded collaborative research project that develops a sustainable AI-Med ecosystem for AI-based medical device development. The project runs from July 2024 to December 2027 and joins seven Northern German research institutions. KiMeKo addresses the complete development trajectory, from concept and data acquisition to validation, regulatory evidence generation, and approval-oriented documentation. The project contributes a practical toolchain and platform capabilities for non-experts and experts, including structured innovation support, uncertaintyaware sensor-data fusion, hybrid expert-system modeling, and workflow-guided data acquisition and anonymization. This paper summarizes project objectives, expected outputs, relevance to IEEE COMPSAC 2026 themes, and current progress. In particular, KiMeKo aligns with Applied AI and Smart & Connected Health by combining AI engineering, privacy-conscious data processing, and regulation-aware medical software development.

This work investigates how self-supervised learning (SSL) frameworks encode semantic structure within their latent representations using the introspection technique Local Concept Embeddings (LoCE).We analyse three complementary SSL paradigms-contrastive (Barlow Twins), generative (Denoising Autoencoder), and predictive (Relative Patch Location)-all pretrained on the Cityscapes dataset and evaluated in a semantic segmentation setting.LoCE reveals that the Denoising Autoencoder produces the most distinct and coherent concept clusters (highest separability and clustering metrics), while Barlow Twins and RPL exhibit moderate structure and higher intra-class variability.Furthermore we find that greater latent disentanglement before fine-tuning correlates with improved segmentation performance, uncovering an interesting link between latent organization and downstream generalization.

Collimation during radiography, which is the process of defining the area to be radiated, is a crucial factor for the protection of the patient and for the diagnostic quality of a radiograph. Moreover, incorrect collimation is one of the main causes for a retake and the associated costs. In this paper we propose a novel collimation optimization approach using depth cameras and deep Neural Networks trained end-to-end. We have acquired two new datasets for this purpose. The first, obtained in a clinical environment, consists of depth images of the lower leg and abdomen and the second, captured in real clinical practice, consists of depth images and corresponding radiographs of thorax examinations. For all depth images, the ideal collimation was labeled by experts either on the depth image or directly on the radiograph. Using this dataset to learn to predict the optimal collimation, we show that it is possible to learn different shapes of collimations and to achieve results that are on par with those obtained by radiographers. Such an AI assistant trained with optimal collimation could reduce the radiation exposure to which the patient is exposed, improve the workflow in radiography, and finally increase the diagnostic quality of radiographs.

Barlow Twins is a feature-contrastive self-supervised learning framework built on the principle of redundancy reduction.The idea is to train a network by maximizing the correlation between corresponding features and minimizing the correlation between non-corresponding features in distorted views of the same image, through this facilitating effective pretraining of a backbone network for a subsequent classification head.This is achieved by diagonalizing the cross-correlation matrix of the network's representations and scaling it towards the identity matrix.We show that the cross-correlation matrix of distorted images is inherently symmetric, independent of the backbone network's weights, which leads to two key insights: (i) the cross-correlation matrix can always be diagonalized using a linear transformation (layer), and (ii) the core idea of maximizing correlations between corresponding features while minimizing them for non-corresponding features alone is insufficient for effective backbone network pretraining.Nevertheless, Barlow Twins provide highly effective pretraining.We show that this is due to the normalization of the cross-correlation matrix in the Barlow Twins cost function.This normalization leads to minima of the cost function which are equivalent to the minima of sample contrastive approaches to enforce invariance.

In clinical practice, a substantial amount of data is generated on a daily basis for diagnostic purposes.Since expensive expert knowledge is required for data annotation in order to use this data for supervised learning, large amounts of data often remain unused.Self-supervised learning methods are well suited for using unlabeled data by pre-training networks to solve pretext tasks.As medical data follow an underlying uneven distribution of occurring diseases, they are inherently imbalanced.This could introduce an unwanted bias during pre-training, ultimately leading to negative consequences that may inhibit the benefits of finetuning.In this work we investigate the impact of the imbalance of 2D and 3D medical datasets used for pre-training, as well as the importance of the type and size of the dataset used for pre-training and the pretext task.Our findings indicate that the size of the dataset used for pre-training has greater impact on the final tasks than its balance.

We study over-parameterized classifiers where empirical risk minimization (ERM) for learning leads to zero training error. In these over-parameterized settings, there are many global minima with zero training error, some of which generalize better than others. We show that under certain conditions, the fraction of "bad" global minima with a true error larger than $\varepsilon $ decays to zero exponentially fast with the number of training data n. The bound depends on the distribution of the true error over the set of classifier functions used for the given classification problem, and does not necessarily depend on the size or complexity (e.g., the number of parameters) of the classifier function set. This insight provides an alternative perspective on the unexpectedly good generalization even of highly over-parameterized neural networks. We substantiate our theoretical findings through experiments on synthetic data and a subset of MNIST. Additionally, we assess our hypothesis using VGG19 and ResNet18 on a subset of Caltech101.

Collimation during radiography, which is the process of defining the area to be radiated, is a crucial factor for the protection of the patient and for the diagnostic quality of a radiograph.Moreover, incorrect collimation is one of the main causes for a retake and the associated costs.In this paper we propose a novel collimation optimization approach using Time-of-Flight cameras and deep Neural Networks trained end-to-end to increase the diagnostic quality of a radiograph.For this we acquired a new dataset in a clinical environment consisting of depth images of the lower leg and the abdomen.Using this dataset we are able to segment depth images for the optimal collimation with an average IoU of 83%.* Contributed equally.The order of author names was randomly determined.† We thank Celina Schubbe for her help and support in collecting the dataset.

Activation maximization (AM) strives to generate optimal input stimuli, revealing features that trigger high responses in trained deep neural networks. AM is an important method of explainable AI. We demonstrate that AM fails to produce optimal input stimuli for simple functions containing ReLUs or Leaky ReLUs, casting doubt on the practical usefulness of AM and the visual interpretation of the generated images. This paper proposes a solution based on using Leaky ReLUs with a high negative slope in the backward pass while keeping the original, usually zero, slope in the forward pass. The approach significantly increases the maxima found by AM. The resulting ProxyGrad algorithm implements a novel optimization technique for neural networks that employs a secondary network as a proxy for gradient computation. This proxy network is designed to have a simpler loss landscape with fewer local maxima than the original network. Our chosen proxy network is an identical copy of the original network, including its weights, with distinct negative slopes in the Leaky ReLUs. Moreover, we show that ProxyGrad can be used to train the weights of Convolutional Neural Networks for classification such that, on some of the tested benchmarks, they outperform traditional networks.

We investigate the learning dynamics of classifiers in scenarios where classes are separable or classifiers are over-parameterized. In both cases, Empirical Risk Minimization (ERM) results in zero training error. However, there are many global minima with a training error of zero, some of which generalize well and some of which do not. We show that in separable classes scenarios the proportion of "bad" global minima diminishes exponentially with the number of training data n. Our analysis provides bounds and learning curves dependent solely on the density distribution of the true error for the given classifier function set, irrespective of the set’s size or complexity (e.g., number of parameters). This observation may shed light on the unexpectedly good generalization of over-parameterized Neural Networks. For the over-parameterized scenario, we propose a model for the density distribution of the true error, yielding learning curves that align with experiments on MNIST and CIFAR-10.

Artificial neural networks oftentimes operate on continuous inputs. While biological neural networks usually represent information through the activity of a population of neurons, the inputs of an artificial neural network are typically provided as a list of scalars. As the information content of each of the input scalars depends heavily on the problem domain, representing them as individual scalar inputs, irrespective of the amount of information they contain, may prove to be suboptimal for the network. We therefore compare and examine four different Population Coding schemes and demonstrate on two toy datasets and one real world benchmark that applying Population Coding to information rich, low dimensional inputs can vastly improve a network’s performance.

We present a novel class of Convolutional Neural Networks called Pre-defined Filter Convolutional Neural Networks (PFCNNs), where all$n\times n$convolution kernels with$n > 1$are pre-defined and constant during training. It involves a special form of depthwise convolution operation called a Pre-defined Filter Module (PFM). In the channel-wise convolution part, the$1\times n\times n$kernels are drawn from a fixed pool of only a few (16) different pre-defined kernels. In the$1\times 1$convolution part linear combinations of the pre-defined filter outputs are learned. Despite this harsh restriction, complex and discriminative features are learned. These findings provide a novel perspective on the way how information is processed within deep CNNs. We discuss various properties of PFCNNs and prove their effectiveness using the popular datasets Caltech101, CIFAR10, CUB-200-2011, FGVC-Aircraft, Flowers102, and Stanford Cars. Our implementation of PFCNNs is provided on Github https://github.com/Criscraft/PredefinedFilterNetworks.

Within the last decade Deep Learning has become a tool for solving challenging problems like image recognition. Still, Convolutional Neural Networks (CNNs) are considered black-boxes, which are difficult to understand by humans. Hence, there is an urge to visualize CNN architectures, their internal processes and what they actually learn. Previously, virtual realityhas been successfully applied to display small CNNs in immersive 3D environments. In this work, we address the problem how to feasibly render large-scale CNNs, thereby enabling the visualization of popular architectures with ten thousands of feature maps and branches in the computational graph in 3D. Our software "DeepVisionVR" enables the user to freely walk through the layered network, pick up and place images, move/scale layers for better readability, perform feature visualization and export the results. We also provide a novel Pytorch module to dynamically link PyTorch with Unity, which gives developers and researchers a convenient interface to visualize their own architectures. The visualization is directly created from the PyTorch class that defines the Pytorch model used for training and testing. This approach allows full access to the network's internals and direct control over what exactly is visualized. In a use-case study, we apply the module to analyze models with different generalization abilities in order to understand how networks memorize images. We train two recent architectures, CovidResNet and CovidDenseNet on the Caltech101 and the SARS-CoV-2 datasets and find that bad generalization is driven by high-frequency features and the susceptibility to specific pixel arrangements, leading to implications for the practical application of CNNs. The code is available on Github https://github.com/Criscraft/DeepVisionVR.

Raven's Progressive Matrices are a benchmark originally designed to test the cognitive abilities of humans. It has recently been adapted to test relational reasoning in machine learning systems. For this purpose the so-called Procedurally Generated Matrices dataset was set up, which is so far one of the most difficult relational reasoning benchmarks. Here we show that deep neural networks are capable of solving this benchmark, reaching an accuracy of 98.0 percent over the previous state-of-the-art of 62.6 percent by combining Wild Relation Networks with Multi-Layer Relation Networks and introducing Magnitude Encoding, an encoding scheme designed for late fusion architectures.

In visual recognition tasks, such as image classification, unsupervised learning exploits cheap unlabeled data and can help to solve these tasks more efficiently. We show that the recursive autoconvolution operator, adopted from physics, boosts existing unsupervised methods by learning more discriminative filters. We take well established convolutional neural networks and train their filters layer-wise. In addition, based on previous works we design a network which extracts more than 600k features per sample, but with the total number of trainable parameters greatly reduced by introducing shared filters in higher layers. We evaluate our networks on the MNIST, CIFAR-10, CIFAR-100 and STL-10 image classification benchmarks and report several state of the art results among other unsupervised methods.

On the way to understanding complex perception tasks based on psychophysical data alone, we propose a general framework using multivariate analysis methods to derive a low-dimensional mapping of the underlying perception space. Psychophysical data can be interpreted in two fundamentally different ways: That is, the characterization of stimuli (e.g. colors) using several verbal descriptions (e.g. bright) - a stimuli-as-points view - and conversely, the characterization of the given verbal descriptions by several stimuli - a descriptors-as-points view. We argue that only the latter view enables us to reach an objective mapping of the perception space. For color perception, we show how it is possible to derive objective maps of the perception space, just based on non-comparative verbal descriptions of color stimuli. We also give an example where we analyze odor perception in the same way to derive a quantitative map of odor perception, a perceptual space that is still fairly unknown in its detailed structure.

Few models exist that accurately reproduce the complex rhythms of the thalamocortical system that are apparent in measured scalp EEG and at the same time, are suitable for large-scale simulations of brain activity. Here, we present a neural mass model of the thalamocortical system during natural non-REM sleep, which is able to generate fast sleep spindles (12-15 Hz), slow oscillations (<1 Hz) and K-complexes, as well as their distinct temporal relations, and response to auditory stimuli. We show that with the inclusion of detailed calcium currents, the thalamic neural mass model is able to generate different firing modes, and validate the model with EEG-data from a recent sleep study in humans, where closed-loop auditory stimulation was applied. The model output relates directly to the EEG, which makes it a useful basis to develop new stimulation protocols.

Recognizing objects in natural images is an intricate problem involving multiple conflicting objectives. Deep convolutional neural networks, trained on large datasets, achieve convincing results and are currently the state-of-the-art approach for this task. However, the long time needed to train such deep networks is a major drawback. We tackled this problem by reusing a previously trained network. For this purpose, we first trained a deep convolutional network on the ILSVRC-12 dataset. We then maintained the learned convolution kernels and only retrained the classification part on different datasets. Using this approach, we achieved an accuracy of 67.68% on CIFAR-100, compared to the previous state-of-the-art result of 65.43%. Furthermore, our findings indicate that convolutional networks are able to learn generic feature extractors that can be used for different tasks.

In this paper we propose the novel algorithm GF-OSC, which learns an orthogonal basis that provides an optimal K-sparse data representation for a given set of training samples. The underlying optimization problem is composed of two nested subproblems: (i) given a basis, to determine an optimal K-sparse coefficient vector for each data sample, and (ii) given a K-sparse coefficient vector for each data sample, to determine an optimal basis. Both subproblems have closed form solutions, which can be computed alternately in an iterative manner. Due to the nesting of the subproblems, however, this approach can only find an optimal solution if the underlying sparsity level is sufficiently high. To overcome this shortcoming, our GF-OSC algorithm solves subproblem (ii) via gradient descent on the corresponding cost function within the underlying lower dimensional space of free dictionary parameters. This algorithmic substep is based on the geodesic flow optimization framework proposed by Plumbley. On synthetic data, we show in a comparison with four alternative learning algorithms the superior recovery performance of GF-OSC and show that it needs significantly fewer learning epochs to converge. Furthermore, we demonstrate the potential of GF-OSC for image compression. For five standard test images, we derived sparse image approximations based on a GF-OSC basis that was trained on natural image patches. In terms of PSNR, the approximation performance of the GF-OSC basis is between 0.09 to 0.32 dB higher compared to using the 2D DCT basis, and between 1.66 to 3.4 dB higher compared to using the 2D Haar wavelet basis.

NREM sleep is characterized by two hallmarks, namely K-complexes (KCs) during sleep stage N2 and cortical slow oscillations (SOs) during sleep stage N3. While the underlying dynamics on the neuronal level is well known and can be easily measured, the resulting behavior on the macroscopic population level remains unclear. On the basis of an extended neural mass model of the cortex, we suggest a new interpretation of the mechanisms responsible for the generation of KCs and SOs. As the cortex transitions from wake to deep sleep, in our model it approaches an oscillatory regime via a Hopf bifurcation. Importantly, there is a canard phenomenon arising from a homoclinic bifurcation, whose orbit determines the shape of large amplitude SOs. A KC corresponds to a single excursion along the homoclinic orbit, while SOs are noise-driven oscillations around a stable focus. The model generates both time series and spectra that strikingly resemble real electroencephalogram data and points out possible differences between the different stages of natural sleep.

By minimizing the zero-norm of the separating hyperplane, the support feature machine (SFM) finds the smallest subspace (the least number of features) of a data set such that within this subspace, two classes are linearly separable without error. This way, the dimensionality of the data is more efficiently reduced than with support vector-based feature selection, which can be shown both theoretically and empirically. In this letter, we first provide a new formulation of the previously introduced concept of the SFM. With this new formulation, classification of unbalanced and nonseparable data is straightforward, which allows using the SFM for feature selection and classification in a large variety of different scenarios. To illustrate how the SFM can be used to identify both the smallest subset of discriminative features and the total number of informative features in biological data sets we apply repetitive feature selection based on the SFM to a functional magnetic resonance imaging data set. We suggest that these capabilities qualify the SFM as a universal method for feature selection, especially for high-dimensional small-sample-size data sets that often occur in biological and medical applications.

Since visual attention-based computer vision applications have gained popularity, ever more complex, biologically inspired models seem to be needed to predict salient locations (or interest points) in naturalistic scenes. In this paper, we explore how far one can go in predicting eye movements by using only basic signal processing, such as image representations derived from efficient coding principles, and machine learning. To this end, we gradually increase the complexity of a model from simple single-scale saliency maps computed on grayscale videos to spatiotemporal multiscale and multispectral representations. Using a large collection of eye movements on high-resolution videos, supervised learning techniques fine-tune the free parameters whose addition is inevitable with increasing complexity. The proposed model, although very simple, demonstrates significant improvement in predicting salient locations in naturalistic videos over four selected baseline models and two distinct data labeling scenarios.

BACKGROUND: Bioluminescence is a process in which light is emitted by a living organism. Most creatures that emit light are sea creatures, but some insects, plants, fungi etc, also emit light. The biotechnological application of bioluminescence has become routine and is considered essential for many medical and general technological advances. Identification of bioluminescent proteins is more challenging due to their poor similarity in sequence. So far, no specific method has been reported to identify bioluminescent proteins from primary sequence. RESULTS: In this paper, we propose a novel predictive method that uses a Support Vector Machine (SVM) and physicochemical properties to predict bioluminescent proteins. BLProt was trained using a dataset consisting of 300 bioluminescent proteins and 300 non-bioluminescent proteins, and evaluated by an independent set of 141 bioluminescent proteins and 18202 non-bioluminescent proteins. To identify the most prominent features, we carried out feature selection with three different filter approaches, ReliefF, infogain, and mRMR. We selected five different feature subsets by decreasing the number of features, and the performance of each feature subset was evaluated. CONCLUSION: BLProt achieves 80% accuracy from training (5 fold cross-validations) and 80.06% accuracy from testing. The performance of BLProt was compared with BLAST and HMM. High prediction accuracy and successful prediction of hypothetical proteins suggests that BLProt can be a useful approach to identify bioluminescent proteins from sequence information, irrespective of their sequence similarity. The BLProt software is available at http://www.inb.uni-luebeck.de/tools-demos/bioluminescent%20protein/BLProt.

BACKGROUND: Results of phylogenetic analysis are often visualized as phylogenetic trees. Such a tree can typically only include up to a few hundred sequences. When more than a few thousand sequences are to be included, analyzing the phylogenetic relationships among them becomes a challenging task. The recent frequent outbreaks of influenza A viruses have resulted in the rapid accumulation of corresponding genome sequences. Currently, there are more than 7500 influenza A virus genomes in the database. There are no efficient ways of representing this huge data set as a whole, thus preventing a further understanding of the diversity of the influenza A virus genome. RESULTS: Here we present a new algorithm, "PhyloMap", which combines ordination, vector quantization, and phylogenetic tree construction to give an elegant representation of a large sequence data set. The use of PhyloMap on influenza A virus genome sequences reveals the phylogenetic relationships of the internal genes that cannot be seen when only a subset of sequences are analyzed. CONCLUSIONS: The application of PhyloMap to influenza A virus genome data shows that it is a robust algorithm for analyzing large sequence data sets. It utilizes the entire data set, minimizes bias, and provides intuitive visualization. PhyloMap is implemented in JAVA, and the source code is freely available at http://www.biochem.uni-luebeck.de/public/software/phylomap.html.

Recently, the so-called Support Feature Machine (SFM) was proposed as a novel approach to feature selection for classification, based on minimisation of the zero norm of a separating hyper plane. We propose an extension for linearly non-separable datasets that allows a direct trade-off between the number of misclassified data points and the number of dimensions. Results on toy examples as well as real-world datasets demonstrate that this method is able to identify relevant features very effectively.

In many industrial applications, Fourier descriptors are commonly used when the description of the object shape is an important characteristic of the image. However, these descriptors are limited to single objects. We propose a general Fourier-based approach, called statistical Fourier descriptor (SFD), which computes shape statistics in grey level images. The SFD is computationally efficient and can be used for defect image classification. In a first example, we deployed the SFD to the inspection of welding seams with promising results.

The well-known MinOver algorithm is a slight modification of the perceptron algorithm and provides the maximum-margin classifier without a bias in linearly separable two-class classification problems. DoubleMinOver as an extension of MinOver, which now includes a bias, is introduced. An O(t(-1)) convergence is shown, where t is the number of learning steps. The computational effort per step increases only linearly with the number of patterns. In its formulation with kernels, selected training patterns have to be stored. A drawback of MinOver and DoubleMinOver is that this set of patterns does not consist of support vectors only. DoubleMaxMinOver, as an extension of DoubleMinOver, overcomes this drawback by selectively forgetting all nonsupport vectors after a finite number of training steps. It is shown how this iterative procedure that is still very similar to the perceptron algorithm can be extended to classification with soft margins and be used for training least squares support vector machines (SVMs). On benchmarks, the SoftDoubleMaxMinOver algorithm achieves the same performance as standard SVM software.

We describe an open-source software framework that simulates the measurements made using one or several cameras in a video-oculographic eye tracker. The framework can be used to compare objectively the performance of different eye tracking setups (number and placement of cameras and light sources) and gaze estimation algorithms. We demonstrate the utility of the framework by using it to compare two remote eye tracking methods, one using a single camera, the other using two cameras.

OneClassMaxMinOver (OMMO) is a simple incremental algorithm for one-class support vector classification. We propose several enhancements and heuristics for improving model selection, including the adaptation of well-known techniques such as kernel caching and the evaluation of the feasibility gap. Furthermore, we provide a framework for optimising grid search based model selection that compromises of preinitialisation, cache reuse, and optimal path selection. Finally, we derive simple heuristics for choosing the optimal grid search path based on common benchmark datasets. In total, the proposed modifications improve the runtime of model selection significantly while they are still simple and adaptable to a wide range of incremental support vector algorithms.

In this paper we address the reliability of policies derived by Reinforcement Learning on a limited amount of observations. This can be done in a principled manner by taking into account the derived Q-functionpsilas uncertainty, which stems from the uncertainty of the estimators used for the MDPpsilas transition probabilities and the reward function. We apply uncertainty propagation parallelly to the Bellman iteration and achieve confidence intervals for the Q-function. In a second step we change the Bellman operator as to achieve a policy guaranteeing the highest minimum performance with a given probability. We demonstrate the functionality of our method on artificial examples and show that, for an important problem class even an enhancement of the expected performance can be obtained. Finally we verify this observation on an application to gas turbine control.

In this brief paper, we propose a method of feature extraction for digit recognition that is inspired by vision research: a sparse-coding strategy and a local maximum operation. We show that our method, despite its simplicity, yields state-of-the-art classification results on a highly competitive digit-recognition benchmark. We first employ the unsupervised Sparsenet algorithm to learn a basis for representing patches of handwritten digit images. We then use this basis to extract local coefficients. In a second step, we apply a local maximum operation to implement local shift invariance. Finally, we train a support vector machine (SVM) on the resulting feature vectors and obtain state-of-the-art classification performance in the digit recognition task defined by the MNIST benchmark. We compare the different classification performances obtained with sparse coding, Gabor wavelets, and principal component analysis (PCA). We conclude that the learning of a sparse representation of local image patches combined with a local maximum operation for feature extraction can significantly improve recognition performance.

We describe an algorithm for manipulating the temporal resolution of a video in real time, contingent upon the viewer's direction of gaze. The purpose of this work is to study the effect that a controlled manipulation of the temporal frequency content in real-world scenes has on eye movements. We build on the work of Perry and Geisler [1998; 2002], who manipulate spatial resolution as a function of gaze direction, allowing them to mimic the resolution distribution of the human retina or to simulate the effect of various diseases (e.g. glaucoma).Our temporal filtering algorithm is similar to that of Perry and Geisler in that we interpolate between the levels of a multiresolution pyramid. However, in our case, the pyramid is built along the temporal dimension, and this requires careful management of the buffering of video frames and of the order in which the filtering operations are performed. On a standard personal computer, the algorithm achieves real-time performance (30 frames per second) on high-resolution videos (960 by 540 pixels).We present experimental results showing that the manipulation performed by the algorithm reduces the number of high-amplitude saccades and can remain unnoticed by the observer.

The well-known MinOver algorithm is a simple modification of the perceptron algorithm and provides the maximum margin classifier in a linearly separable two class classification problem. In its dual formulation selected training patterns which determine the separating hyperplane have to be stored. A drawback of MinOver is that this set of patterns does not consist only of support vectors. With MaxMinOver an extension of MinOver by a simple forgetting procedure is introduced. It is shown that this forgetting not only reduces the number of patterns which have to be stored, but also improves convergence bounds. After a finite number of training steps, the set of stored training patterns will consist only of support vectors. It is shown how this simple and iterative procedure can also be extended to classification with soft margins. The SoftMaxMinOver algorithm exhibits close connections to the v/support-vector-machine.

Detecting the sites on genomic DNA at which DNA binding proteins bind is a highly relevant task in bioinformatics. For example, the binding sites of transcription factors are key elements of regulatory networks and determine the location of genes on a genome. Usually, for a given DNA binding protein, only a few DNA-subsequences at which the protein binds are known experimentally. The task then is to deduce the global binding characteristics of the protein based on these few positive examples. A widespread approach is the so-called profile-matrix (PM). The PM-approach can be interpreted as a linear classifier (binding word class/non-binding word class) within the space of sequence words, with the profile of the experimentally verified binding sites determining its parameters. In this paper a novel approach called binding-matrix (BM) is introduced. Like the PM, the BM realizes a linear classification, but in contrast to the profile-matrix approach the parameters (matrix) of the classifier is now determined by maximum likelihood estimation. Tested on data from the TRANSFAC database, the maximum likelihood estimation leads to an increase in classification performance by about an order of magnitude.

The neighborhood preservation of self-organizing feature maps like the Kohonen map is an important property which is exploited in many applications. However, if a dimensional conflict arises this property is lost. Various qualitative and quantitative approaches are known for measuring the degree of topology preservation. They are based on using the locations of the synaptic weight vectors. These approaches, however, may fail in case of nonlinear data manifolds. To overcome this problem, in this paper we present an approach which uses what we call the induced receptive fields for determining the degree of topology preservation. We first introduce a precise definition of topology preservation and then propose a tool for measuring it, the topographic function. The topographic function vanishes if and only if the map is topology preserving. We demonstrate the power of this tool for various examples of data manifolds.

A neural network algorithm based on a soft-max adaptation rule is presented. This algorithm exhibits good performance in reaching the optimum minimization of a cost function for vector quantization data compression. The soft-max rule employed is an extension of the standard K-means clustering procedure and takes into account a neighborhood ranking of the reference (weight) vectors. It is shown that the dynamics of the reference (weight) vectors during the input-driven adaptation procedure are determined by the gradient of an energy function whose shape can be modulated through a neighborhood determining parameter and resemble the dynamics of Brownian particles moving in a potential determined by the data point density. The network is used to represent the attractor of the Mackey-Glass equation and to predict the Mackey-Glass time series, with additional local linear mappings for generating output values. The results obtained for the time-series prediction compare favorably with the results achieved by backpropagation and radial basis function networks.

A hierarchical neural network structure capable of learning the control of a robot's arm and gripper is introduced. Based on T. Kohonen's algorithm (1982) for the formation of topologically correct feature maps and on an extension of the algorithm for learning of output signals, a simulated robot arm system learns the task of grasping a cylinder. The network architecture is that of a 3-D cubic lattice in which is nested at each lattice node a 2-D square lattice. The robot learns without supervision to position its arm and to orient its gripper properly by observing its own trial movements. In a simulation, the error in positioning the manipulator after training was 0.3% of the robot's dimension, and the residual error in orienting the gripper was 3.8°. Due to cooperation between neighboring neurons during the training phase, fewer than two trial movements per neuron were sufficient to learn the required control tasks

An extension of T. Kohonen's (1982) self-organizing mapping algorithm together with an error-correction scheme based on the Widrow-Hoff learning rule is applied to develop a learning algorithm for the visuomotor coordination of a simulated robot arm. Learning occurs by a sequence of trial movements without the need for an external teacher. Using input signals from a pair of cameras, the closed robot arm system is able to reduce its positioning error to about 0.3% of the linear dimensions of its work space. This is achieved by choosing the connectivity of a three-dimensional lattice consisting of the units of the neural net.