Branko Šter

Designable DNA-binding domains enable construction of logic circuits in mammalian cells

Electronic computer circuits consisting of a large number of connected logic gates of the same type, such as NOR, can be easily fabricated and can implement any logic function. In contrast, designed genetic circuits must employ orthogonal information mediators owing to free diffusion within the cell. Combinatorial diversity and orthogonality can be provided by designable DNA- binding domains. Here, we employed the transcription activator-like repressors to optimize the construction of orthogonal functionally complete NOR gates to construct logic circuits. We used transient transfection to implement all 16 two-input logic functions from combinations of the same type of NOR gates within mammalian cells. Additionally, we present a genetic logic circuit where one input is used to select between an AND and OR function to process the data input using the same circuit. This demonstrates the potential of designable modular transcription factors for the construction of complex biological information-processing devices.

An integrated learning approach to environment modelling in mobile robot navigation

Abstract We extend the approach to learning a topological description of the environment with recurrent neural networks. Usually, a predetermined reactive behavior and a predefined criterion for decision points are used. In our extended approach, both the reactive behavior and the criterion for the decision points are adaptive and therefore more flexible. The reactive behavior is learnt using reinforcement learning supplemented by a new, psychologically grounded mechanism that enables the robot to autonomously explore the environment in a useful way for the purposes of modelling. Decision points or situations where a deviation from the reactive behavior is allowed are learnt on-line using a novel criterion based on the information theory. Results of experiments conducted with a simulated mobile robot equipped with proximity sensors and a color video camera show applicability of the proposed approach.

Asymmetric clustering using the alpha-beta divergence

We propose the use of an asymmetric dissimilarity measure in centroid-based clustering. The dissimilarity employed is the Alpha-Beta divergence (AB-divergence), which can be asymmetrized using its parameters. We compute the degree of asymmetry of the AB-divergence on the basis of the within-cluster variances. In this way, the proposed approach is able to flexibly model even clusters with significantly different variances. Consequently, this method overcomes one of the major drawbacks of the standard symmetric centroid-based clustering.

Markov random field model for segmenting large populations of lipid vesicles from micrographs.

Giant unilamellar lipid vesicles, artificial replacements for cell membranes, are a promising tool for in vitro assessment of interactions between products of nanotechnologies and biological membranes. However, the effect of nanoparticles can not be derived from observations on a single specimen, vesicle populations should be observed instead. We propose an adaptation of the Markov random field image segmentation model which allows detection and segmentation of numerous vesicles in micrographs. The reliability of this model with different lighting, blur, and noise characteristics of micrographs is examined and discussed. Moreover, the automatic segmentation is tested on micrographs with thousands of vesicles and the result is compared to that of manual segmentation. The segmentation step presented is part of a methodology we are developing for bio-nano interaction assessment studies on lipid vesicles.

Prediction of intended career choice in family medicine using artificial neural networks

Abstract Background: Due to the importance of family medicine and a relative shortage of doctors in this discipline, it is important to know how the decision to choose a career in this field is made. Objective: Since this decision is closely linked to students’ attitudes towards family medicine, we were interested in identifying those attitudes that predict intended career choice in family medicine. Methods: A cross-sectional study was performed among 316 final-year medical students of the Ljubljana Medical Faculty in Slovenia. The students filled out a 164-item questionnaire, developed based on the European definition of family medicine and the EURACT Educational Agenda, using a seven-point Likert scale containing attitudes towards family medicine. The students also recorded their interest in family medicine on a five-point Likert scale. Attitudes were selected using a feature selection procedure with artificial neural networks that best differentiated between students who are likely and students who are unlikely to become family physicians. Results: Thirty-one out of 164 attitudes predict a career in family medicine, with a classification accuracy of at least 85%. Predictors of intended career choice in family medicine are related to three categories: understanding of the discipline, working in a coherent health care system and person-centredness. The most important predictor is an appreciation of a long-term doctor–patient relationship. Conclusion: Students whose intended career choice is family medicine differ from other students in having more positive attitudes towards family physicians’ competences and towards characteristics of family medicine and primary care.

Adaptive Radial Basis Decomposition by Learning Vector Quantization

A method for function approximation in reinforcement learning settings is proposed. The action-value function of the Q-learning method is approximated by the radial basis function neural network and learned by the gradient descent. Those radial basis units that are unable to fit the local action-value function exactly enough are decomposed into new units with smaller widths. The local temporal-difference error is modelled by a two-class learning vector quantization algorithm, which approximates distributions of the positive and of the negative error and provides the centers of the new units. This method is especially convenient in cases of smooth value functions with large local variation in certain parts of the state space, such that non-uniform placement of basis functions is required. In comparison with four related methods, it has the smallest requirements of basis functions when achieving a comparable accuracy.

Enhancing upper confidence bounds for trees with temporal difference values

Upper confidence bounds for trees (UCT) is one of the most popular and generally effective Monte Carlo tree search (MCTS) algorithms. However, in practice it is relatively weak when not aided by additional enhancements. Improving its performance without reducing generality is a current research challenge. We introduce a new domain-independent UCT enhancement based on the theory of reinforcement learning. Our approach estimates state values in the UCT tree by employing temporal difference (TD) learning, which is known to outperform plain Monte Carlo sampling in certain domains. We present three adaptations of the TD(λ) algorithm to the UCTs tree policy and backpropagation step. Evaluations on four games (Gomoku, Hex, Connect Four, and Tic Tac Toe) reveal that our approach increases UCTs level of play comparably to the rapid action value estimation (RAVE) enhancement. Furthermore, it proves highly compatible with a modified all moves as first heuristic, where it considerably outperforms RAVE. The findings suggest that integration of TD learning into MCTS deserves further research, which may form a new class of MCTS enhancements.

Final year medical students’ understanding of family medicine

OBJECTIVE The European Academy of Teachers in General Practice / Family Medicine (EURACT) has developed an educational agenda, the key document for teaching family medicine in Europe. The aim of our study was to find out how final year medical students at the beginning of their family medicine clerkship understand the discipline of family medicine. METHODS The attitudes toward family medicine were paraphrased and developed into a 164-item questionnaire, which was administered to 335 final-year medical students at the beginning of their clerkship. Using combinatorial optimization with genetic algorithms we selected 30 items which yielded the highest Cronbach alpha reliability coefficient. Finally, we performed a factor analysis to find which dimensions of family medicine were recognised by the students and compared them with the domains defined in the EURACT definition. RESULTS The 30-item questionnaire had a Cronbach alpha reliability coefficient of 0.919. The differences between male and female students were not very significant (p=0.061). With the factor analysis we recognised seven factors, belonging to three out of six domains of the EURACT educational agenda: primary care management, personcenteredness and comprehensive approach. CONCLUSION Final-year medical students at the beginning of their family medicine clerkship understand some of the dimensions of family medicine rather well, but they are not aware of some important competences of family doctors. There is a necessity to teach students about specific problem solving skills and the importance of balance between the health needs of an individual patient and the community.

Structural Properties of Recurrent Neural Networks

In this article we research the impact of the adaptive learning process of recurrent neural networks (RNN) on the structural properties of the derived graphs. A trained fully connected RNN can be converted to a graph by defining edges between pairs od nodes having significant weights. We measured structural properties of the derived graphs, such as characteristic path lengths, clustering coefficients and degree distributions. The results imply that a trained RNN has significantly larger clustering coefficient than a random network with a comparable connectivity. Besides, the degree distributions show existence of nodes with a large degree or hubs, typical for scale-free networks. We also show analytically and experimentally that this type of degree distribution has increased entropy.

Impact of Learning on the Structural Properties of Neural Networks

We research the impact of the learning process of neural networks (NN) on the structural properties of the derived graphs. A type of recurrent neural network is used (GARNN). A graph is derived from a NN by defining a connection between any pair od nodes having weights in both directions above a certain threshold. We measured structural properties of graphs such as characteristic path lengths (L), clustering coefficients (C) and degree distributions (P). We found that well trained networks differ from badly trained ones in both Land C.