Color coding
(HCI) Application
Computer Experiments
Machine Learning
Math
Neuroscience
Software

2006

14. Efficiency of goal-oriented communicating agents in different graph topologies: A study with Internet crawlers
A. Lőrincz, K.A. Lázár, and Zs. Palotai
Physica A, 1doi:10.1016/j.physa.2006.11.052, 2006
Abstract
To what extent does the communication make a goal-oriented community efficient in different topologies? In order to gain insight into this problem, we study the influence of learning method as well as that of the topology of the environment on the communication efficiency of crawlers in quest of novel information in different topics on the Internet. Individual crawlers employ selective learning, function approximation-based reinforcement learning (RL), and their combination. Selective learning, in effect, modifies the starting URL lists of the crawlers, whilst RL alters the URL orderings. Real data have been collected from the web and scale-free worlds, scale-free small world (SFSW), and random world environments (RWEs) have been created by link reorganization. In our previous experiments [Zs. Palotai, Cs. Farkas, A. Lőrincz, Is selection optimal in scale-free small worlds?, ComPlexUs 3 (2006) 158–168], the crawlers searched for novel, genuine documents and direct communication was not possible. Herein, our finding is reproduced: selective learning performs the best and RL the worst in SFSW, whereas the combined, i.e., selective learning coupled with RL is the best—by a slight margin—in scale-free worlds. This effect is demonstrated to be more pronounced when the crawlers search for different topic-specific documents: the relative performance of the combined learning algorithm improves in all worlds, i.e., in SFSW, in SFW, and in RWE. If the tasks are more complex and the work sharing is enforced by the environment then the combined learning algorithm becomes at least equal, even superior to both the selective and the RL algorithms in most cases, irrespective of the efficiency of communication. Furthermore, communication improves the performance by a large margin and adaptive communication is advantageous in the majority of the cases..

13. Social network analysis: Measuring tools, structures and dynamics (Editorial)
A. Lőrincz, N. Gilbert, and R. Goolsby
Physica A : doi: 10.1016/j.physa.2006.11.042, 2006

12. Neural Network Solution for Real and Complex Independent Subspace Analysis
Z. Szabó and A. Lőrincz
ICA Research Network Workshop , ICA Research Network International Workshop, Sept 18-19, Liverpool, UK 1: 85-88, 2006
Abstract
Uncovering independent processes is of high importance, because it breaks combinatorial explosion (Póczos and Lőrincz, 2006). In cases, like Smart Dust, the problem is vital, because (i) elements have limited computational capacity and (ii) communication to remote distances is prohibitively expensive. Self-adjusting, self-grouping neural network solutions may come to our rescue here. Here, we present such an approach for Independent Subspace Analysis (ISA). Our method (i) generalizes (Meyer-Bäse et al., 2006), (ii) can be used both for real and complex processes, and (iii) can be related to kernel based techniques [3, 4]. The extension of ISA to Independent Process Analysis is straightforward under certain conditions (Póczos and Lőrincz, 2006).

11. Simple conditions for forming triangular grids
B. Takács and A. Lőrincz
Neurocomputing (in press)
Abstract
We have used simple learning rules to study how firing maps containing triangular grids -- as found in \emph{in vivo} experiments -- can be developed by Hebbian means in realistic robotic simulations. We started from typical non-local postrhinal neuronal responses. We found that anti-Hebbian weight pruning can develop triangular grids under certain conditions. Experimental evidences and the present study suggest that within this model, whitening is a bottom-up process, whereas weight pruning and possibly the non-linear extension of whitening to bottom-up information maximization are guided by top-down influences that reorganize entorhinal responses. We connect our model to the computational model of the entorhinal-hippocampal region of Lőrincz and Buzsáki. In the joined model, the hippocampus is the origin of response reorganization. The joined model may provide insights for memory reorganization guided by hippocampal supervision.

10. Critical Echo State Networks
M.A. Hajnal and A. Lőrincz
Lecture Notes in Computer Science 4131: 658-667, 2006
Abstract
We are interested in the optimization of the recurrent connection structure of Echo State Networks (ESNs). It is known that the topology of these connections can strongly influence performance. We study the predictive capacity by numerical simulations on Mackey-Glass time series, and find that a particular small subset of ESNs is much better than ordinary ESNs provided that the topology of the recurrent feedback connections satisfies certain conditions. We argue that the small subset separates two large sets of ESNs and this separation can be characterized in terms of phase transitions. With regard to the criticality of this phase transition, we introduce the notion of Critical Echo State Networks (CESN). We discuss why CESNs perform better than other ESNs.

9. Reinforcement Learning with Echo State Networks
I. Szita, V. Gyenes and A. Lőrincz
Lecture Notes in Computer Science 4131: 830-839, 2006
Abstract
Function approximators are often used in reinforcement learning tasks with large or continuous state spaces. Artificial neural networks, among them recurrent neural networks are popular function approximators, especially in tasks where some kind of of memory is needed, like in real-world partially observable scenarios. However, convergence guarantees for such methods are rarely available. Here, we propose a method using a class of RNNs recently proposed, called echo state networks. Proof of convergence to a bounded region is provided for k-order Markov decision processes. Runs on illustrative POMDP problems were performed to test and illustrate the working of the architecture.

8. Emerging artificial societies through learning
N. Gilbert, M. den Besten, Á. Bontovics, B.G.W. Craenen, F. Divina, A.E. Eiben, R. Griffioen, Gy. Hévízi, A. Lőrincz, B. Paechter, S. Schuster, M.C. Schut, C. Tzolov, P. Vogt and L. Yang
Journal of Artificial Societies and Social Simulation 9, 2006
URL: http://jasss.soc.surrey.ac.uk/9/2/9.html
Abstract
The NewTies project is implementing a simulation in which societies of agents are expected to develop autonomously as a result of individual, population and social learning. These societies are expected to be able to solve environmental challenges by acting collectively. The challenges are intended to be analogous to those faced by early, simple, small-scale human societies. This report on work in progress outlines the major features of the system as it is currently conceived within the project, including the design of the agents, the environment, the mechanism for the evolution of language and the peer-to-peer infrastructure on which the simulation runs.

7. A Framework for Anonymous but Accountable Self-Organizing Communities
G. Ziegler, Cs. Farkas and A. Lőrincz
Information and Software Technology 48: 726-744, 2006
Abstract
In this paper we propose a novel approach to provide accountability for Web communities that require a high-level of privacy.We target Web communities that are dynamical and selforganizing, i.e., roles and contributions of participants may change over time. We present a framework that supports dynamic grouping and collaboration. Our security protocols build upon a semi-trusted computing base and community-based trust. A two-layered privacy protection architecture is proposed, that supports (i) evaluation of participants and enforcement of community rules, called internal accountability and (ii) rule-based interaction with real world organizations, called external accountability. The two layered architecture limits the exposure of the users' con.dential information, such as the mappings between real users, their virtual identities, and among the virtual users, while keeps the potential to ful- .ll the legal requirements imposed by the environment. Our concepts and protocols are implemented in our SyllabNet project that supports university students to evaluate courses.

6. Computer study of the evolution of 'news foragers' on the Internet
Zs. Palotai, S Mandusitz, and A. Lőrincz
In.: Swarm Intelligence and Data Mining, Springer SCI Series, 34: 203-220, Springer, Berlin, Germany, 2006
Eds.: A. Abraham, C. Grosan, and V. Ramos
Abstract
We populated a huge scale-free portion of Internet environment with news foragers. They evolved by a simple internal selective algorithm: selection concerned the memory components, being finite in size and containing the list of most promising supplies. Foragers received reward for locating not yet found news and crawled by using value estimation. Foragers were allowed to multiply if they passed a given productivity threshold. A particular property of this community is that there is no direct interaction (here, communication) amongst foragers that allowed us to study compartmentalization, assumed to be important for scalability, in a very clear form. Experiments were conducted with our novel scalable Alife architecture. These experiments had two particular features. The first feature concerned the environment: a scale-free world was studied as the space of evolutionary algorithms. The choice of this environment was due to its generality in mother nature. The other feature of the experiments concerned the fitness. Fitness was not predetermined by us, but it was implicitly determined by the unknown, unpredictable environment that \textit{sustained} the community and by the evolution of the competitive individuals. We found that the Alife community achieved fast compartmentalization.

5. Non-combinatorial Estimation of Independent Autoregressive Sources
B. Póczos and A. Lőrincz
Neurocomputing 69: 2416-2419, 2006
Abstract
Identification of mixed independent subspaces is thought to suffer from combinatorial explosion of two kinds: the minimization of mutual information between the estimated subspaces and the search for the optimal number and dimensions of the subspaces. Here we show that independent auto-regressive process analysis, under certain conditions, can avoid this problem using a two-phase estimation process. We illustrate the solution by computer demonstration.

4. Learning Tetris Using the Noisy Cross-Entropy Method
I. Szita and A. Lőrincz
Neural Computation 18: 2936-2941, 2006
Abstract
The cross-entropy method is an efficient and general optimization algorithm. However, its applicability in reinforcement learning seems to be limited although it is fast, because it often converges to suboptimal policies. A standard technique for preventing early convergence is to introduce noise. We apply the noisy cross-entropy method to the game of Tetris to demonstrate its efficiency. The resulting policy outperforms previous RL algorithms by almost two orders of magnitude, and reaches nearly 300,000 points on average.

3. Cross-Entropy Optimization for Independent Process Analysis
Z. Szabó, B. Póczos and A. Lőrincz
Lecture Notes in Computer Science 3889:909-916, Springer-Verlag, 2006
Abstract
We treat the problem of searching for hidden multi-dimensional independent auto-regressive processes. First, we transform the problem to independent subspace analysis (ISA). Our main contribution concerns ISA. We show that under certain conditions, ISA is equivalent to a combinatorial optimization problem. For the solution of this optimization we apply the cross-entropy method. Numerical studies indicate that the cross-entropy method can offer an order of magnitude gain in precision over other state-of-the-art methods.

2. Epsilon-Sparse Representations: Generalized sparse approximation and the equivalent family of SVM tasks
Z. Szabó and A. Lőrincz
Acta Cybernetica 17(3): 605-614, 2006 (Acta Cybernetica reference page)
Abstract
Relation between a family of generalized Support Vector Machine (SVM) problems and the novel epsilon-sparse representation is provided. In defining epsilon-sparse representations, we use a natural generalization of the classical epsilon-insensitive cost function for vectors. The insensitive parameter of the SVM problem is transformed into component-wise insensitivity and thus overall sparsification is replaced by component-wise sparsification. The connection between these two problems is built through the generalized Moore-Penrose inverse of the Gram matrix associated to the kernel. The relevance of the epsilon-sparse representation is mentioned.

1. Independent component analysis forms place cells in realistic robot simulations
B. Takács and A. Lőrincz
Neurocomputing 69: 1249-1252, 2006
Abstract
It has been argued that the processing of sensory information in the entorhinal-hippocampal loop involves independent component analysis (ICA) on temporally concatenated inputs. Here, we demonstrate that ICA in a realistic robot simulation on a U-shaped track forms place fields similar to those found in rat experiments in vivo.

2005

20. Independent Kalman-filter model of the entorhinal-hippocampal loop
A. Lőrincz
Society for Neuroscience Meeting, November 2005
Abstract
The recent model of the EC-HC loop (Lorincz and Buzsaki, NYAS 911:83-111, 2000) is generalized as follows: The EC-HC loop is a prototype of self-structuring generative neural assemblies that enable independent modulations. This conjecture can be unfolded: (1) The system is made of parallel processing independently adjustable overcomplete neuronal assemblies, (2) internal representations of the assemblies are selected and optimized by their predictive generative capabilities subject to the actual context, (3) computations, learning and denoising are local. The model that satisfies these constraints works as follows: The EC-HC loop estimates and forms independent Kalman filters (IKFs) that can be (a) controlled top-down and (b) selected by temporal contexts, making the network highly nonlinear. Blind learning occurs in the HC. Loops formed by the dentate gyrus (DG), mossy cells and CA3 subfield execute blind source deconvolution. Thus, the output of the CA3 suits blind source separation (BSS or ICA). ICA maps CA3 to CA1. Kalman-gain is regulated locally by inhibitory networks at the CA1. ICA is also a preprocessing step for independent subspace analysis (ISA). ISA forms the subspaces of the IKFs. IKFs make the columns of the EC. EC layer 6 contains the hidden representation of the IKFs. Recurrent collaterals of EC layer 6 form predictive networks that learn to predict the hidden process and form the infinite impulse response (IIR) filter of the IKF model. Deep layers optimize the IKF. Long-term memory includes excitatory, deep to superficial feedback connections in the EC and also between cortical areas. LTM formation requires two-phase operation. Layer 2 is a comparator that outputs the IIR error of the IKFs to the DG and to the CA3 subfield. Errors computed by the EC-HC loop support self-supervised learning, including the linear embedding of the non-linear echo state neural network formed by layers 2 and 3 that learn and can reconstruct deterministic portions of the input. Layer 3 is modulated by layer 5. EC layer 3 efferents to the CA1 perform ICA and pave the way for bottom-up information. Hebbian learning arises in many cases.

19. PIRANHA: Policy Iteration for Recurrent Artificial neural Networks with Hidden Activities
I. Szita and A. Lőrincz
Neurocomputing 70: 577-591, 2006.
Abstract
It is an intriguing task to develop efficient connectionist representations for learning long time series. Recurrent neural networks have great promises here. We model the learning task as a minimization problem of a nonlinear least-squares cost function, that takes into account both one-step and multi-step prediction errors. The special structure of the cost function is constructed to build a bridge to reinforcement learning. We exploit this connection and derive a convergent, policy iteration-based algorithm, and show that RNN training can be made to fit the reinforcement learning framework in a natural fashion. The relevance of this connection is discussed. We also present experimental results, which demonstrate the appealing properties of the unique parameter set prescribed by reinforcement learning. Experiments cover both sequence learning and long-term prediction.

18. Is selection optimal for scale-free small worlds?
Zs. Palotai, Cs. Farkas and A. Lőrincz
European Conference on Complex Systems (ECCS) 2005 (accepted for long talk)
ComPlexUs 3: 158-168, 2006
Abstract
The 'No Free Lunch Theorem' claims that for the set of all problems no algorithm is better than random search. Thus, selection can be advantageous only on a limited set of problems. We investigated how the topological structure of the environment influences algorithmic efficiency. We have studied random, scale-free, and scale-free small world (SFSW) topologies. Selective learning, reinforcement learning and their combinations were tried. We found that selective learning is the best in SFSW topology. In non-small world topologies, however, selection looses against the combined algorithm. Learning agents were web-crawlers searching for novel, not-yet-found information. Experiments were run on a large news site and on its downloaded portion. Controlled experiments were performed on this downloaded portion: we modified the topology, but kept the occurrence times of the news. Our findings may have implications for the theory of evolution

17. Investigating Complexity with the New Ties Agent
A.R. Griffioen, Á. Bontovics, A.E. Eiben, Gy. Hévízi, and A. Lőrincz
European Conference on Complex Systems (ECCS) 2005 (poster)
Abstract

16. The New Ties project: 3 dimensions of adaptivity and 3 dimensions of complexity scale-up
A.E. Eiben, N. Gilbert, A. Lőrincz, B. Paechter, P. Vogt
European Conference on Complex Systems (ECCS) 2005 (poster)
Abstract

15. A Corpus-Based Neural Net Method for Explaining Unknown Words by WordNet Senses
B. Gábor, V. Gyenes and A. Lőrincz
Proc. of. ECML/PKDD 2005 Lecture Notes in Artificial Intelligence 3721: 470-477 Springer-Verlag
Abstract
This paper introduces an unsupervised algorithm that collects senses contained in WordNet to explain words, whose meaning is unknown, but plenty of documents are available that contain the word in the desired sense. Based on the widely accepted idea that the meaning of a word is characterized by its context, a reconstruction network approach was designed to reconstruct the meaning of the unknown word. The connections of the network were derived from word co-occurrences and word-sense statistics. The approach was found robust against details of the architecture. The method was tested on 80 TOEFL synonym questions, from which the system was able to answer 61 correctly. This is better than the average performance of non-native students applying to universities in the United States, and is comparable to other methods tested on the same questions, but using a larger corpus or richer lexical database.

14. Independent Subspace Analysis on Innovations
B. Póczos, B. Takács and A. Lőrincz
Proc. of. ECML/PKDD 2005 Lecture Notes in Artificial Intelligence 3720: 698-706, Springer-Verlag
Abstract
Independent subspace analysis (ISA), where the independent sources can be multi-dimensional, is a generalization of independent component analysis (ICA). However, to our best knowledge, all known ISA algorithms are ineffective when the sources possess temporal structure. Hyvärinen proposed the innovation process instead of the original mixtures to solve ICA problems with temporal dependencies. Here we show that employing the innovation process extends the range of problems for ISA as well. We demonstrate the idea on an illustrative 3D problem and also on mixtures of facial pictures as two-dimensional sources. When samples were taken in pixelwise order, ISA on innovations was able to find the original subspaces, while plain ISA was not.

13. Independent subspace analysis using geodesic spanning trees
B. Póczos and A. Lőrincz
Proc. of Int. Conf. on Machine Learing, Bonn 2005 ICML: 673-680 (2005)
Abstract
A novel algorithm for performing Independent Subspace Analysis, the estimation of hidden independent subspaces is introduced. This task is a generalization of Independent Component Analysis. The algorithm works by estimating the multi-dimensional differential entropy. The estimation utilizes minimal geodesic spanning trees matched to the sample points. Numerical studies include (i) illustrative examples, (ii) a generalization of the cocktail-party problem to songs played by bands, and (iii) an example on mixed independent subspaces, where subspaces have dependent sources, which are pairwise independent.

12. Independent subspace analysis using k-nearest neighborhood distances
B. Póczos and A. Lőrincz
Proc. of ICANN 2005 Lecture Notes in Computer Science 3697: 163-168, Springer-Verlag, 2005
Abstract
A novel algorithm called independent subspace analysis (ISA) is introduced for estimating independent subspaces. The algorithm solves the ISA problem by estimating multi-dimensional differential entropies. Two variants are examined, both of them utilize distances between the $k$-nearest neighbors of the sample points. Numerical simulations demonstrate the usefulness of the algorithms.

11. Selection in scale-free small world
Zs. Palotai, Cs. Farkas, and A. Lőrincz
Proc of. CEEMAS 2005 Lecture Notes in Artificial Intelligence 3690: 579-582, Springer-Verlag
Abstract
In this paper we compare our selection based learning algo-rithm with the reinforcement learning algorithm in Web crawlers. The task of the crawlers is to ¯nd new information on the Web. The selection algorithm, called weblog update, modi¯es the starting URL lists of our crawlers based on the found relevant documents. The reinforcement learning algorithm modi¯es the URL orderings of the crawlers based on the received reinforcements for submitted documents. We performed simulations based on data collected from the Web. The collected portion of the Web is typical and exhibits scale-free small world (SFSW) structure. We have found that on this SFSW, the weblog update algorithm performs better than the reinforcement learning algorithm. It ¯nds the new information faster than the reinforcement learning algorithm and has better new information/all submitted documents ratio.

10. Assisting robotic personal agent and cooperating alternative input devices for severely disabled children
Gy. Hévízi, B. Gerőfi, B. Szendrő, and A. Lőrincz
Proc. of CEEMAS 2005, Lecture Notes in Artificial Intelligence 3690: 591-594, Springer-Verlag
Abstract
In this paper we present a multi-component cooperating system made of alternative input devices, software tools and special instruments designed for severely disabled children having various disabilities. Different input tools have been developed to exploit possible 'outputs', e.g., head motion or leg motion that might be controlled or might learned to be controlled by these children. Speci.c software tools serve to convert these 'outputs' in di.erent computer aided tasks. Simple, but extensible software communication framework ensures the con.gurable connection between them. The goal is to iterate with caretakers and to develop a software, which is manageable by the parents, who can select and adjust the applicable components to the capabilities of the children. The system also enables the permanent monitoring of user activities with possible future bene.ts for evaluations. A speci.c arrangement of these tools is detailed here. This arrangement incorporates RF-MEMS tools, a head motion monitoring webcam, and a robotic personal agent. These tools together, enable communication with people not in the immediate neighborhood.

9. Embedded body sensor network for persons with special communication needs to control and to interact with the world — DEMO
A. Lőrincz, Gy. Hévízi, Zs. Palotai, and B. Szendrő
Proceedings of the International Workshop on Wearable and Implantable Body Sensor Networks April 12-13, 2005 (Imperial College, London)
IEE London, ISBN 0-86341-505-9. p. 117. (2005).
BSN 2005 home
Abstract
There is an ongoing project at the Eötvös Loránd University. We have been developing tools for severely handicapped, non-speaking but speech understanding children. Main components of the project are (i) the development of special communication tools including a controllable pet, Aibo, (ii) the development of machine learning techniques for situational optimization, and (iii) the development of Body Sensor Network (BSN) to form ambient intelligence.
Type 1: Controlled operation. In this case, Aibo is controlled by the user. Aibo's head follows the head movements of the user, and the images from Aibo's camera are transmitted to the user. Aibo speaks instead of the user and transmits sound to the user.
Type 2: Assisting operation. Aibo is not controlled by the user and it is watching anyone in the neighborhood. It monitors head and gaze directions and transfers information to the user, according to the assumed context suggested by the user's head direction. During interaction, user performance can be used for RL.
Type 3: Interacting operation. Aibo is not controlled by the user, it is watching the user. It is not transmitting received signals, it is in reactive mode. It reacts to gaze and head directions of the user and tries to interact with user. During this interaction, attention, smile, unhappiness, anger and other signs can be used for RL
Download extended abstract (pdf)

8. ICA forms place cells, Hebbian learning forms replaying recurrent collaterals in robotic simulations
B. Takács and A. Lőrincz
CNS 2005 (accepted)
Abstract
It has been argued that the processing of sensory information in the loop of the hippocampus and the entorhinal cortex involves independent component analysis (ICA) on temporally concatenated inputs. We demonstrate that ICA applied to a realistic Khepera robot simulation can form receptive fields found \emph{in vivo} in rat experiments. Novel and simple Hebbian form emerges for CA3 collaterals, which learn to predict. We have found that this `add-on' learning spares receptive field properties. We conjecture that the entorhinal afferents of the CA1 subfield learn to denoise \textit{and also} provide supervisory information for the Schaffer collaterals, thus keeping the CA1 representation steady during the slower formation of the bottom-up ICA and the predictive collateral connection system of the CA3 subfield. This model accounts for recent findings on the CA3 and CA1 place cells.

7. Competitive spiking and indirect entropy minimization of rate code: Efficient search for hidden components
B. Szatmáry, B. Póczos and A. Lőrincz
Journal of Physiology (Paris) 98: 407-416 (2004).
Abstract
Our motivation, which originates from the psychological and physiological evidences of component-based representations in the brain, is to find neural methods that can efficiently search for structures. Here, an architecture made of coupled parallel working reconstruction subnetworks is presented. Each subnetwork utilizes non-negativity constraint on the generative weights and on the internal representation. `Spikes' are generated within subnetworks via winner take all mechanism. Memory components are modified by a noisy stochastic gradient method that minimizes the reconstruction error. Learning rate is subject to changes. The dynamics of the learning rate is ruled by a novel indirect entropy minimization procedure: the higher the entropic interpretation of the rate code, the higher the learning rate in the subnetworks. This method effectively reduces the search space and increases the escape probability from high entropy local minima. We demonstrate coupled networks can discover and sort components into the subnetworks; a problem subject to combinatorial explosion. Synergy between spike code and rate code is discussed.

6. NEW TIES Agent
A.R. Griffioen, M.C. Schut, A.G. Eiben, Á. Bontovics, Gy. Hévízi and A. Lőrincz
AISB'05 Convention, Hatfield UK, April 12-14, 2005
Proc. of the Socially Inspired Computing Joint Symposium, pp. 72-79, ISBN 1 902956 48 4
Abstract
In this paper we discuss issues relating to modelling autonomous agents. The agents are situated in a realistic environment where they have to survive for extended periods of time by means of learning. We present among others the architecture of the agent and propose an algorithmic solution to integrate evolutionary and lifetime learning.

5. Selection in scale-free small world
Zs. Palotai, Cs. Farkas and A. Lőrincz
(submitted)
Abstract
In this paper we compare Web crawlers using our selection based algorithm and the known to be well performing reinforcement learning algorithm to forage new information on a news web site. The selection algorithm, called weblog update, modifies the starting URL lists of our crawlers. The reinforcement learning algorithm modifies the URL orderings of the crawlers. We performed realistic simulations based on data collected during a crawl on the Web. The collected portion of the Web is typical and exhibits scale-free small world (SFSW) structure. We have found that on this SFSW, the weblog update algorithm performs better than the reinforcement learning algorithm. It finds the new information faster than the reinforcement learning algorithm and has better new information/all submitted documents ratio.We argue that the clear advantages of selection over reinforcement learning may be limited to small worlds.

4. Headmouse and machine learning for persons with special communication needs
B. Takács, T. Keller-Márkus, S.T. Dakio, Gy. Hévízi, B. Póczos, Z. Szabó, S.L. Kálmán and A. Lőrincz
(submitted)
Abstract
Cursor driven by head motion has several advantages for people who need AAC. Here, an affordable headmouse system (HMS) is presented. The system has been tested for normal participants and for persons with special communication needs. For normal participants, we have tested the usefulness of our HMS as a writing tool. For persons with special communication needs our objective was to examine if they can improve the control of their head motion to sufficient precision. The HMS seems excellent for this purpose, because of the direct visual feedback that this tool provides. Our results are promising in both respects. We also show that the computer can learn to cluster user behaviors that may enable adaptive user assistance in the future.

3. Adaptive highlighting of links to assist surfing on the Internet
Zs. Palotai, B. Gábor and A. Lőrincz
International Journal of Information Technology & Decision Making 4: 117-139 (2005)
Abstract
The largest source of information is the WWW. Gathering of novel information from this network constitutes a real challenge for artificial intelligence (AI) methods. Large search engines do not offer a satisfactory solution, their indexing cycle is long and creates a time lag of about one month. Moreover, sometimes search engines offer a huge amount of documents, which is hard to constrain and to increase the ratio of relevant information. A novel AI-assisted surfing method, which highlights links during surfing is studied here. The method makes use of (i) `experts', i.e. pre-trained classifiers, forming the long-term memory of the system, (ii) relative values of experts and value estimation of documents based on recent choices of the users. Value estimation adapts fast and forms the short-term memory of the system. (iii) Neighboring documents are downloaded, their values are estimated and valuable links are highlighted. Efficiency of the idea is tested on an artificially generated sample set, on a downloaded portion of the Internet and in real Internet searches using different models of the user. All experiments show that surfing based filtering can efficiently highlight 10-20% of the documents in about 5 to 10 steps, or less.

2. Attentional filtering in neocortical areas: A top-down model
A. Lőrincz
Neurocomputing 65: 817-823, 2005.
CNS 2004, Baltimore, Collection of Abstracts (http://www.cnsorg.org/Abstract_Book.pdf) No. 341 (T22) p. 108 (2004).
Abstract
Two comparator based rate code models -- a reconstruction network model and a control model -- are merged. The role of bottom-up filtering is information maximization and noise filtering, whereas top-down control `paves the way' of context based information prediction that we consider as attentional filtering. Falsifying prediction of the model has gained experimental support recently.

1. Neural Kalman-filter
G. Szirtes, B. Póczos and A. Lőrincz
Neurocomputing 65: 349-355, 2005.
CNS 2004, Baltimore, Collection of Abstracts (http://www.cnsorg.org/Abstract_Book.pdf) No. 245 (M27) p. 84 (2004).
Abstract
Anticipating future events is a crucial function of the central nervous system and can be modelled by Kalman-filter like mechanisms which are optimal for predicting linear dynamical systems. Connectionist representation of such mechanisms with Hebbian learning rules has not yet been derived. We show that the recursive prediction error method offers a solution that can be mapped onto the entorhinal-hippocampal loop in a biologically plausible way. Model predictions are provided.

2004

14. Noise induced structures in STDP networks
G. Szirtes, Zs. Palotai and A. Lőrincz
CNS 2004, Baltimore, Collection of Abstracts (http://www.cnsorg.org/Abstract_Book.pdf) No. 272 (M53) p. 93 (2004).
Abstract
In this paper we study the emergent structures in networks with spike-timing dependent synaptic plasticity that are subject to external noise. We show that scale-free small-worlds can emerge in such noise driven Hebbian networks. This may explain the interplay between noise and Hebbian plasticity. We also argue that this model can be seen as a unification of the Watts-Strogatz and preferential attachment models of scale-free small-worlds.

13. Is neocortical encoding of sensory information intelligent?
A. Lőrincz
CNS 2004, Baltimore, Collection of Abstracts (http://www.cnsorg.org/Abstract_Book.pdf) No. 151 (S31) p. 61 (2004).
Abstract
The theory of computational complexity is used to underpin a recent model of neocortical sensory processing. It is argued that the theory of computational complexity points to generative networks and that these networks resolve the homunculus fallacy. Computational simulations illustrate the idea.

12. Finding structure by entropy minimization in coupled reconstruction networks
B. Szatmáry, B. Póczos and A. Lőrincz
CNS 2004, Baltimore, July 18-20, Collection of Abstracts (http://www.cnsorg.org/Abstract_Book.pdf) No. 124 (S05) p. 44 (2004).
Abstract
There is psychological and physiological evidence for components-based representations in the brain. We present a special architecture of coupled parallel working reconstruction subnetworks that can learn components of input and extract the structure of these components. Each subnetwork directly minimizes the reconstruction error and indirectly minimizes the entropy of the internal representation via a novel tuning method, which effectively reduces the search space by changing the learning rate dynamically and increasing the escape probability from local minima. Revealing the structure of the input improves when competitive spiking and indirect minimization of the entropy of spike rate are applied together.

11. Intelligent encoding and economical communication in the visual stream
A. Lőrincz
Early Cognitive Vision Workshop, Workshop on Coding of Visual Information in the Brain, June 1, 2004, Isle of Skye, Scotland
http://cogprints.ecs.soton.ac.uk/archive/00003505/, http://arxiv.org/pdf/q-bio.NC/0403022,
http://www.cn.stir.ac.uk/ecovision-ws/pdf/71.pdf.
Abstract
The theory of computational complexity is used to underpin a recent model of neocortical sensory processing. We argue that encoding into reconstruction networks is appealing for communicating agents using Hebbian learning and working on hard combinatorial problems, which are easy to verify. Computational definition of the concept of intelligence is provided. Simulations illustrate the idea.

10. Bottom-up clustering and top-down shattering of scale-free environments for information fusion
A. Lőrincz, B. Gábor, S. Mandusitz and Zs. Palotai
Information Fusion 2004, June 28 - July 1, 2004, Stockholm, Sweden
http://www.afosr.af.mil/pages/AFOSRWorkshop.htm, http://www.fusion2004.foi.se/proc.html,
http://www.fusion2004.foi.se/papers/IF04-0471.pdf.
Abstract
Scale-free small world graphs, such as the Internet, have nodes of high clustering coefficients and are ideally suited for clustering. Our goal is to enable context-sensitive characterization of Internet documents. Four algorithms are being developed and are under testing in various Internet environments. The weblog algorithm utilizes competitive agents and shatters the Internet by collecting topic specific nodes of high clustering coeffi- cients from different domains. The bottom-up clustering algorithm develops a tree-structured and easy to visualize cluster hierarchy. Keyword extracting algorithm chooses the best keywords that match a subset of the clusters. Link-highlighting models user reinforcement and ranks Internet documents during navigation. The architecture of these algorithms allows for user reinforcement and iterated clustering. The aim of the joined architecture is to adapt the representation to the user’s requests.

9. Simple algorithm for recurrent neural networks that can learn sequence completion
I. Szita and A. Lőrincz
Int. Joint Conf. on Neural Networks 2004, 26 - 29 July 2004, Budapest, Paper No. 1101. IJCNN2004 CD-ROM Conference Proceedings, IEEE Catalog Number: 04CH37541C, ISBN: 0-7803-8360-5, IEEE Operations Center, Piscataway, NJ 08855-1331.
Abstract
We can memorize long sequences like melodies or poems and it is intriguing to develop efficient connectionist representations for this problem. Recurrent neural networks have been shown to offer a reasonable approach here. We start from a few axiomatic assumptions and provide a simple mathematical framework that encapsulates the problem. A gradient-descent based algorithm is derived in this framework. Demonstrations on a benchmark problem show the applicability of our approach.

8. Distributed novel news mining from the Internet with an evolutionary news forager community
Zs. Palotai, S. Mandusitz and A. Lőrincz
Int. Joint Conf. on Neural Networks 2004, 26 - 29 July 2004, Budapest, Paper No. 1095. IJCNN2004 CD-ROM Conference Proceedings, IEEE Catalog Number: 04CH37541C, ISBN: 0-7803-8360-5, IEEE Operations Center, Piscataway, NJ 08855-1331.
Abstract
We populated a huge scale-free portion of Internet environment with news foragers. They evolved by a simple internal selective algorithm: selection concerned the memory components, being finite in size and containing the list of most promising supplies. Foragers received reward for locating not yet found news and crawled by using value estimation. Foragers were allowed to multiply if they passed a given productivity threshold. A particular property of this community is that there is no direct interaction (here, communication) amongst foragers. It is found that, still, fast compartmentalization, i.e., fast division of work can be achieved, enabling us to comment on symbiotic dynamical hierarchies.

7. Hidden Markov model finds behavioral patterns of users working with a headmouse driven writing tool
Gy. Hévízi, M. Biczó, B. Póczos, Z. Szabó, B. Takács and A. Lőrincz
Int. Joint Conf. on Neural Networks 2004, 26 - 29 July 2004, Budapest, Paper No. 1268. IJCNN2004 CD-ROM Conference Proceedings, IEEE Catalog Number: 04CH37541C, ISBN: 0-7803-8360-5, IEEE Operations Center, Piscataway, NJ 08855-1331.
Abstract
We studied user behaviors when the cursor is directed by a head in a simple control task. We used an intelligent writing tool called Dasher. Hidden Markov models (HMMs) were applied to separate behavioral patterns. We found that a similar interpretations can be given to the hidden states upon learning. It is argued that the recognition of such general application specific behavioral patterns should be of help for adaptive humancomputer interfaces.

6. Emerging evolutionary features in noise driven STDP networks?
Zs. Palotai, G. Szirtes and A. Lőrincz
Int. Joint Conf. on Neural Networks 2004, 26 - 29 July 2004, Budapest, Paper No. 1119. IJCNN2004 CD-ROM Conference Proceedings, IEEE Catalog Number: 04CH37541C, ISBN: 0-7803-8360-5, IEEE Operations Center, Piscataway, NJ 08855-1331.
Abstract
In this paper we study the emergent structure of networks in which spike-timing dependent synaptic plasticity is induced only by external random noise. We show that such noise driven Hebbian networks are able to develop a broad range of network structures, including scale-free small-world networks. The development of such network structures may provide an explanation of the role of noise and its interplay with Hebbian plasticity. We also argue that this model can be seen as a unification of the famous Watts-Strogatz and preferential attachment models of small-world and scale-free nets. Our results may support Edelman’s idea on that the development of central nervous system may have evolutionary components.

5. Value estimation based computer-assisted data mining for surfing the Internet
B. Gábor, Zs. Palotai and A. Lőrincz
Int. Joint Conf. on Neural Networks 2004, 26 - 29 July 2004, Budapest, Paper No. 1035. IJCNN2004 CD-ROM Conference Proceedings, IEEE Catalog Number: 04CH37541C, ISBN: 0-7803-8360-5, IEEE Operations Center, Piscataway, NJ 08855-1331.
Abstract
Gathering of novel information from the WWW constitutes a real challenge for artificial intelligence (AI) methods. Large search engines do not offer a satisfactory solution, their indexing cycle is long and they may offer a huge amount of documents. An AI-based link-highlighting procedure designed to assist surfing is studied here. It makes use of (i) `experts', i.e. pre-trained classifiers, forming the long-term memory of the system, (ii) relative values of experts and value estimation of documents based on recent choices of the users. Value estimation adapts fast and forms the short-term memory of the system. All experiments show that surfing based filtering can efficiently highlight 10-20% of the documents in about 5 steps, or less.

4. An algorithm for finding reliably schedulable plans
B. Takács, I. Szita and A. Lőrincz
Int. Joint Conf. on Neural Networks 2004, 26 - 29 July 2004, Budapest, Paper No. 1098. IJCNN2004 CD-ROM Conference Proceedings, IEEE Catalog Number: 04CH37541C, ISBN: 0-7803-8360-5, IEEE Operations Center, Piscataway, NJ 08855-1331.
Abstract
For interacting agents in time-critical applications, learning whether a subtask can be scheduled reliably is an important issue. The identification of sub-problems of this nature may promote e.g. planning, scheduling and segmenting in Markov decision processes. We define a subtask to be schedulable if its execution time has a small variance. We present an algorithm for finding such subtasks.

3. Improving recognition accuracy on structured documents by learning structural patterns
Gy. Hévízi, T. Marcinkovics and A. Lőrincz
Pattern Analysis and Applications 7: 66-76 (2004)
Abstract
In this paper we present an `add-on' probabilistic method that can improve the efficiency of known document classification algorithms when applied to structured documents. This method is designed to augment other, known but non-probabilistic classification schemes of structured documents or information extraction procedures to reduce uncertainties. To this end a probabilistic distribution on the structure of XML documents is introduced. It shown how to parameterize existing learning methods to describe the structure distribution efficiently. The learned distribution is then used to predict the classes of unseen documents. Novelty detection making use of the structure-based distribution function is also discussed. Demonstration on model documents and on Internet XML documents are presented.

2. Temporal plannability by variance of the episode length
B. Takács, I. Szita and A. Lőrincz
Technical Report
Abstract
For interacting agents in time-critical applications, learning of the possibility of scheduling subtasks is an important issue. The identification of sub-problems of this nature may promote e.g. planning, scheduling and segmenting Markov decision processes. Formulae for the standard deviation of the duration are derived and illustrated.

1. Kalman filter control embedded into the reinforcement learning framework
I. Szita and A. Lőrincz
Neural Computation 16: 491-499, 2004.
Abstract
There is a growing interest in using Kalman filter models in brain modeling. In turn, the question arises whether Kalman filter models can be used on-line not only for estimation but control. The usual method of optimal control of Kalman filter makes use of off-line backward recursion, which is not satisfactory for this purpose. Here, it is shown that a slight modification of the linear-quadratic-Gaussian Kalman filter model allows the on-line estimation of optimal control by using reinforcement learning and overcomes this difficulty. Moreover, the emerging learning rule for value estimation exhibits a Hebbian form, which is weighted by the error of the value estimation.

2003

8. Reinforcement learning with linear function approximation and LQ control converges
I. Szita and A. Lőrincz
Technical Report
Supplement to paper in Neural Computation16: 491-499, 2004.
Abstract
Reinforcement learning is commonly used with function approximation. However, very few positive results are known about the convergence of function approximation based RL control algorithms. In this paper we show that TD(0) and Sarsa(0) with linear function approximation is convergent for a simple class of problems, where the system is linear and the costs are quadratic (the LQ control problem). Furthermore, we show that for systems with Gaussian noise and non-completely observable states (the LQG problem), the mentioned RL algorithms are still convergent, if they are combined with Kalman filtering.

7. Cost component analysis
A. Lőrincz and B. Póczos
International Journal of Neural Systems 13:, 183-192, 2003.
Abstract
In optimizations the dimension of the problem may severely, sometimes exponentially increase optimization time. Parametric function approximatiors (FAPPs) have been suggested to overcome this problem. Here, a novel FAPP, cost component analysis (CCA) is described. In CCA, the search space is resampled according to the Boltzmann distribution generated by the energy landscape. That is, CCA converts the optimization problem to density estimation. Structure of the induced density is searched by independent component analysis (ICA). The advantage of CCA is that each independent ICA component can be optimized separately. In turn, (i) CCA intends to partition the original problem into subproblems and (ii) separating (partitioning) the original optimization problem into subproblems may serve interpretation. Most importantly, (iii) CCA may give rise to high gains in optimization time. Numerical simulations illustrate the working of the algorithm.

6. Kalman-filtering using local interactions
B. Póczos and A. Lőrincz
Technical Report
Abstract
There is a growing interest in using Kalman-filter models for brain modelling. In turn, it is of considerable importance to represent Kalman-filter in connectionist forms with local Hebbian learning rules. To our best knowledge, Kalman-filter has not been given such local representation. It seems that the main obstacle is the dynamic adaptation of the Kalman-gain. Here, a connectionist representation is presented, which is derived by means of the recursive prediction error method. We show that this method gives rise to attractive local learning rules and can adapt the Kalman-gain.

5. Modeling the 'homunculus'
A. Lőrincz
NEWS of the European Research Consortium for Informatics and Mathematics (ERCIM NEWS), Special Issue on Cognitive Systems53:21-22, 2003.
http://www.ercim.org/publication/Ercim_News/
Abstract
The Neural Information Processing Group of the Eötvös Loránd University Budapest has been engaged in researching reflexive systems, capable to collect experiences, including reinforcement, to learn from those experiences and to accept directions. The envisioned system keeps an eye on itself and may seem that it knows what it is doing.

4. Robust hierarchical image representation using non-negative matrix factorization with sparse code shrinkage preprocessing
B. Szatmáry, G. Szirtes, A. Lőrincz, J. Eggert and E. Körner
Pattern Analysis and Application Journal 6: 194-200 (2003).
Abstract
When analyzing patterns, our goals are (i) to find structure in the presence of noise, (ii) to decompose the observed structure into sub-components, and (iii) to use the components for pattern completion. Here, a novel loop architecture is introduced to perform these tasks in an unsupervised manner. The architecture combines sparse code shrinkage with non-negative matrix factorization and blends their favorable properties: Sparse code shrinkage aims to remove Gaussian noise in a robust fashion; Non-negative matrix factorization extracts sub-structures from the noise filtered inputs. The loop architecture performs robust pattern completion when organized into a two-layered hierarchy. We demonstrate the power of the proposed architecture on the so-called 'bar-problem' and on the Feret facial database.

3. A2SOC: Anonymity and accountability in self-organizing electronic communities
Cs. Farkas, G. Ziegler, A. Meretei and A. Lőrincz
ACM Workshop on Privacy in Electronic Society Nov 18-22, 2002, Washington, D.C., USA: Eds.: S. De Capitani di Vimercati and P. Samarati pp. 81-91 (2003, ACM)
Abstract
In this paper we study the problem of anonymity versus accountability in electronic communities. We argue that full anonymity may present a security risk that is unacceptable in certain applications. Therefore, anonymity and accountability are both needed. To resolve the inherent contradiction between anonymity and accountability in a flexible manner, we introduce the concepts of internal and external accountabilities. Intuitively, internal accountability applies to virtual users only, and is governed by the policy of a group (a community). In contrast, external accountability is needed to address issues related to misuse if the activity is to be penalized in real life according to internal rules or external laws. We provide a set of protocols to ensure that users' virtual and real identities cannot be disclosed unnecessarily, and allow users to monitor the data collected about them as well as to terminate their membership (both real and virtual) under certain conditions. We develop a general conceptual model of electronic Editorial Board (e-EB). In our thinking, there are deep connections between anonymity and self-organization. In turn, the concept of self-organizing e-EB (SO-eEB) is introduced here, and a robotic example is provided. Finally, SO-eEB is specialized to \textit{Anonymous and Accountable Self-Organizing Communities A2SOCs),} that fully supports internal and external accountability while providing anonymity.

2. Event-learning and robust policy heuristics
A. Lőrincz, I. Pólik and I. Szita
Cognitive Systems Research 4: 319-337, 2003.
Abstract
In this paper we introduce a novel reinforcement learning algorithm called event-learning. The algorithm uses \emph{events}, ordered pairs of two consecutive states. We define event-value function and we derive learning rules. Combining our method with a well-known robust control method, the SDS algorithm, we introduce Robust Policy Heuristics (RPH). It is shown that RPH, a fast-adapting non-Markovian policy, is particularly useful for coarse models of the environment and could be useful for some partially observed systems. RPH may be of help in alleviating the `curse of dimensionality' problem. Event-learning and RPH can be used to separate time scales of learning of value functions and adaptation. We argue that the definition of modules is straightforward for event-learning and event-learning makes planning feasible in the RL framework. Computer simulations of a rotational inverted pendulum with coarse discretization are shown to demonstrate the principle.

1. Epsilon-MDPs: Learning in varying environments
I. Szita, B. Takács and A. Lőrincz
Journal of Machine Learning Research3:145-174, 2003.
Abstract
In this paper epsilon-MDP models are introduced and convergence theorems are proven using the generalized MDP framework of Szepesvári and Littman. Using this model family, we show that Q-learning is capable of finding near-optimal policies in varying environments. The potentials of this new family of MDP models are illustrated via a reinforcement learning algorithm called event-learning which separates the optimization of decision making from the controller. We show that event-learning augmented by a particular controller, which gives rise to an epsilon-MDP, enables near optimal performance even if considerable and sudden changes may occur in the environment. Illustrations are provided on the two-segment pendulum problem.

2002

11. HebbNets: Dynamic network with Hebbian learning rule
G. Szirtes, Zs. Palotai and A. Lőrincz
Technical Report
Abstract
It has been demonstrated that one of the most striking features of the nervous system, the so called 'plasticity' (i.e high adaptability at different structural levels) is primarily based on Hebbian learning which is a collection of slightly different mechanisms that modify the synaptic connections between neurons. The changes depend on neural activity and assign a special dynamic behavior to the neural networks. From a structural point of view, it is an open question what network structures may emerge in such dynamic structures under 'sustained' conditions when input to the system is only noise. In this paper we present and study the `HebbNets', networks with random noise input, in which structural changes are exclusively governed by neurobiologically inspired Hebbian learning rules. We show that Hebbian learning is able to develop a broad range of network structures, including scale-free small-world networks.

10. Mystery of structure and function of sensory processing areas of the neocortex: A resolution
A. Lőrincz, B. Szatmáry and G. Szirtes
Journal of Computational Neuroscience 13: 187–205, 2002.
Abstract
Many different neural models have been proposed to account for major characteristics of the memory phenomenon family in primates. However, in spite of the large body of neurophysiological, anatomical and behavioral data, there is no direct evidence for supporting one model while falsifying the others. And yet, we can discriminate models based on their complexity and/or their predictive power. In this paper we present a computational framework with our basic assumption that neural information processing is performed by generative networks. A complex architecture is `derived' by using information-theoretic principles. We find that our approach seems to uncover possible relations among the functional memory units (declarative and implicit memory) and the process of information encoding in primates. The architecture can also be related to the entorhinal-hippocampal loop. An effort is made to form a prototype of this computational architecture and to map it onto the functional units of the neocortex. This mapping leads us to claim that one may gain a better understanding by considering that anatomical and functional layers of the cortex differ. Philosophical consequences regarding the \textit{homunculus fallacy} are also considered.

9. Reinforcement learning integrated with a non-Markovian controller
I. Szita, B. Takács and A. Lőrincz
ECAI2002, Proceedings of the 15th European Conference on Artificial Intelligence, In F. van Harmelen (ed.): IOS Press, Amsterdam, 2002, pp.365-369.
Abstract
Recently a novel reinforcement learning algorithm called event-learning or E-learning was introduced. The algorithm based on events, which are defined as ordered pairs of states. In this setting, the agent optimizes the selection of desired sub-goals by a traditional value-policy function iteration, and utilizes a separated algorithm called the controller to achieve these goals. The advantage of event-learning lies in its potential in non-stationary environments, where the near-optimality of the value iteration is guaranteed by the generalized epsilon-stationary MDP model. Using a particular non-Markovian controller, the SDS controller, an epsilon-stationary MDP problem arises in E-learning. We illustrate the properties of E-learning augmented by the SDS controller by computer simulations.

8. Non-negative matrix factorization extended by sparse code shrinkage and by weight sparsification 
B. Szatmáry, B. Póczos, J. Eggert, E. Körner, A. Lőrincz
ECAI2002, Proceedings of the 15th European Conference on Artificial Intelligence, In F. van Harmelen (ed.): IOS Press, Amsterdam, 2002, pp.503-507.
Abstract
Properties of a novel algorithm called non-negative matrix factorization (NMF), are studied. NMF can discover substructures and can provide estimations about the presence or the absence of those, being attractive for completion of missing information. We have studied the working and learning capabilities of NMF networks. Performance was improved by adding sparse code shrinkage (SCS) algorithm to remove structureless noise. We have found that NMF performance is considerably improved by SCS noise filtering. For improving noise resistance in the learning phase, weight sparsification was studied; a sparsifying prior was applied on the NMF weight matrix. Learning capability versus noise content was measured with and without sparsifying prior. In accordance with observation made by others on independent component analysis, we have also found that weight sparsification improved learning capabilities in the presence of Gaussian noise.

7. Fast adapting value estimation based hybrid architecture for searching the world-wide web
I. Kókai and A. Lőrincz
Applied Soft Computing 2: 11-23, 2002.
Abstract
The slogan that information is power has undergone a slight change. Today, information updating is in the focus of interest. The largest source of information is the world-wide web. Fast search methods are in need for this enormous source. In this paper a hybrid architecture that combines soft support vector classification and reinforcement learning for value estimation is introduced for the evaluation of a link (a document) and its neighboring links (or documents), called the context of a document. The method is motivated by (i) large differences between such contexts on the web, (ii) the facilitation of goal oriented search using context classifiers, and (iii) attractive fast adaptation properties, that could counteract diversity of web environments. We demonstrate that value estimation based fast adaptation offers considerable improvement over other known search methods.

6. Relating priming and repetition suppression
A. Lőrincz, G. Szirtes, B. Takács, I. Biederman, and R. Vogels
International Journal of Neural Systems 12 : 187-202, 2002.
Abstract
We present a prototype of a recently proposed two stage model of the entorhinal-hippocampal loop. Our aim is to form a general computational model of the sensory neocortex. The model -- grounded on pure information theoretic principles -- accounts for the most characteristic features of long-term memory (LTM), performs bottom-up novelty detection, and supports noise filtering. Noise filtering can also serve to correct temporal ordering of information processing. Surprisingly, as we examine the temporal characteristics of the model, the emergent dynamics can be interpreted as perceptual priming, a fundamental type of implicit memory. In the model's framework, computational results support the hypothesis of a strong correlation between perceptual priming and repetition suppression and this correlation is a direct consequence of the temporal ordering in forming the LTM. We also argue that our prototype offers a relatively simple and coherent explanation of priming and its relation to a general model of information processing by the brain.

5. Ockham's razor at work: Modeling of the 'homunculus'
A. Lőrincz, B. Póczos, G. Szirtes, and B. Takács
Brain and Mind 3: 187-220, 2002
Abstract
There is a broad consensus about the fundamental role of the hippocampal system (hippocampus and its adjacent areas) in the encoding and retrieval of episodic memories. This paper presents a functional model of this system. Although memory is not a single-unit cognitive function, we took the view that the whole system of the smooth, interrelated memory processes may have a common basis. That is why we follow the Ockham's razor principle and minimize the size or complexity of our model assumption set. The fundamental assumption is the requirement of solving the so called ``homunculus fallacy'', which addresses the issue of interpreting the input. Generative autoassociators seem to offer a resolution of the paradox. Learning to represent and to recall information, in these generative networks, imply maximization of information transfer, sparse representation and novelty recognition. A connectionist architecture, which integrates these aspects as model constraints, is derived. Numerical studies demonstrate the novelty recognition and noise filtering properties of the architecture. Finally, we conclude that the derived connectionist architecture can be related to the neurobiological substrate.

4. Low level priming as a consequence of perception
G. Szirtes and A. Lőrincz
Connectionist Models of Cognition and Perception, Proc. of the 7th Neural Computation Workshop,  Eds.: J.A. Bullinaria and W. Lowe, World Scientific, Singapore, NCPW7: 223-235, 2002.
Abstract
In this paper we examine the core of a recently proposed model of the entorhinal-hippocampal loop (EHL). The core model is built on pure information theoretic principles. It accounts for the most characteristic features of the EHL including long-term memory (LTM) formation. Here we argue that the core model, which performs novelty detection, provides correct temporal ordering for learning. Surprisingly, as we examine the temporal characteristics of the model, the experienced dynamics can be interpreted as the perceptual priming phenomenon. Computational results support the hypothesis that there might be a strong correlation between perceptual priming and repetition suppression and this correlation is a direct consequence of the temporal ordering in forming the LTM. We argue also that a relatively simple and coherent explanation of priming and its relation to the general model of the brain's information processing system positions our model as a possible prototype for neocortical layers.

3. Categories, prototypes and memory systems in Alzheimer's disease
Sz. Kéri, Z. Janka, Gy. Benedek, P. Aszalós, B. Szatmáry, G. Szirtes, and A. Lőrincz
Trends in Cognitive Science 6: 132-136, 2002.
Abstract
Recent studies indicate that category learning is mediated by multiple neuronal systems. It has been shown that patients with marked impairments in executive functions, explicit memory, and procedural learning can categorize the exemplars and prototype of a previously trained category. Simple, self-organizing neuronal networks can explain prototype learning and related dysfunction in Alzheimer's disease and provide a model of how prototype learning is mediated by circumscribed mechanisms in the visual cortex.

2. Intelligent high-performance crawlers used to reveal topic-specific structure of the WWW
A. Lőrincz, I. Kókai, and A. Meretei
International Journal of Foundations of Computer Science 13: 477-495, 2002
Abstract
The slogan that `information is power' has undergone a slight change. Today, `information updating' is in the focus of interest. The largest source of information today is the World Wide Web. Fast search methods are needed to utilize this enormous source of information. In this paper our novel crawler using support vector classification and on-line reinforcement learning is described. We launched crawler searches from different sites, including sites that offer, at best, very limited information about the search subject. This case may correspond to typical searches of non-experts. Results indicate that the considerable performance improvement of our crawler over other known crawlers is due to its on-line adaptation property. We used our crawler to characterize basic topic-specific properties of WWW environments. It was found that topic-specific regions have a broad distribution of valuable documents. Expert sites are excellent starting points, whereas mailing lists can form traps for the crawler. These properties of the WWW and the emergence of intelligent `high-performance' crawlers that monitor and search for novel information together predict a significant increase of communication load on the WWW in the near future.

1. Value prediction in HLS allocation problems using intellectual properties
Zs. Palotai, T. Kandár, Z. Mohr, T. Visegrády, G. Ziegler, P. Arató, and A. Lőrincz
Applied Artificial Intelligence 16: 151-192, 2002.
Abstract
Value approximation based global search algorithm is suggested to solve resource constrained allocation in high level synthesis problems. Value approximation is preferred, because it can start by using expert heuristics, can estimate the global structure of the search problem and can optimize heuristics. We are concerned by those allocation problems that have hidden global structure that value approximation may unravel. The value approximation applied here computes the cost of the actual solution \emph{and} estimates the cost of the solution that could be achieved upon performing a global search on the hidden structure starting from the actual solution. We transcribed the allocation problem into a special form of weighted CNF formulae to suit our approach. We also extended the formalism to pipeline operations. Comparisons are made with expert heuristics. Scaling of computation time and performance are compared.

2001

9. Independent component analysis of temporal sequences subject to constraints by LGN inputs yields all the three major cell types of the primary visual cortex
B. Szatmáry and A. Lőrincz
Journal of Computational Neuroscience 11: 241-248, 2001.
Abstract
Information maximization has long been suggested as the underlying coding strategy of the primary visual cortex (V1). Grouping image sequences into blocks has been shown by others to improve agreement between experiments and theory. We have studied the effect of temporal convolution on the formation of spatiotemporal filters, i.e. the analogues of receptive fields, since this temporal feature is characteristic to the response function of lagged and non-lagged cells of the LGN. Concatenated input sequences were used to \textit{learn} the linear transformation that maximizes the information transfer. Learning was accomplished by means of principal component analysis and independent component analysis. Properties of the emerging spatiotemporal filters closely resemble the three major types of V1 cells: (i) simple cells with separable receptive field, (ii) the simple cells with non-separable receptive field, (iii) the complex cells.

8. Global ambient intelligence: The hostess concept
A. Lőrincz
NEWS of the European Research Consortium for Informatics and Mathematics (ERCIM NEWS), Special Issue on Ambient Intelligence 47:24-25, 2001. http://www.ercim.org/publication/Ercim_News/
Abstract
The Neural Information Processing Group of the Eötvös Loránd University, Budapest, launched a project 24 months ago to develop a general methodology for distributed collaborating goal-oriented experts (including humans) without using any assumption on synchronization. The expert community will have adaptive means for distributed computing, editing, and decision-making. Medical applications and home monitoring are in focus.

7. Value prediction in engineering applications
G. Ziegler, Z. Palotai, T. Cinkler, P. Arató, and A. Lőrincz
Lecture Notes in Artificial Intelligence, 2070: 25-34, 2001.
Abstract
One of the most fundamental activities in engineering design is global optimisation. In many such cases no eÆecient algorithm is known to get exact results, therefore heuristics are widely used. We have found that function approximator based estimation of the value function is advantageous in two such engineering applications: in Dense Wavelength Division Multiplexing and in High Level Synthesis. In both cases we have rewritten the orignal optimisation task into a suitable formulation for a recent heuristic called STAGE, which fells into the class of adaptive multiple-retarts algorithms. STAGE takes a user de.ned local search algorithm as its \plug-in" and it utilises value estimation for a \smart guess" of the most promising restart point for the local search. It is important that if a global structure of local minima is found by the function approximator then search time may not have to scale with the dimension of the problem in the exponent, but it may become a polynomial function of the dimension.

6. Event learning and robust policy heuristics
A. Lőrincz, I. Pólik and I. Szita
Technical Report NIPG-ELU-14-05-2001 Abstract (html), download pdf, zipped, or postscript version
Abstract
In this paper we introduce a novel form of reinforcement learning called event-learning or E-learning. In our method an event is an ordered pair of two consecutive states. We define event-value function and derive learning rules which are guaranteed to converge to the optimal event-value function. Combining our method with a known robust control method, the SDS algorithm, we introduce Robust Policy Heuristics (RPH). It is shown that RPH, a fast-adapting non-Markovian policy, is particularly useful for coarse models of the environment and for partially observed systems. As such, RPH alleviates the `curse of dimensionality' problem. Fast adaptation can be used to separate time scales of learning the value functions of a Markovian decision making problem and adaptation, the utilization of a non-Markovian policy. We shall argue that (i) the definition of modules is straightforward for E-learning, (ii) E-learning extends naturally to policy switching, and (iii) E-learning promotes planning. Computer simulations of a two-link pendulum with coarse discretization and noisy controller are shown to demonstrate the principle.

5. Independent component analysis of temporal sequences forms place cells
A. Lőrincz, G. Szirtes, B. Takács, and Gy. Buzsáki
Neurocomputing 38: 769-774, 2001
Abstract
It has been suggested that sensory information processing makes use of a factorial code. It has been shown that the major components of the hippocampal-entorhinal loop can be derived by conjecturing that the task of this loop is forming and encoding independent components (ICs), one type of factorial codes. However, continuously changing environmen poses additional requirements on the coding that can be (partially) satisfied by extendi time slices. We use computer simulations to decide whether IC analysis on temporal sequences can produce place fields in labyrinths or not.

4. Sign-changing filters similar to cells in primary visual cortex emerge by independent component analysis of temporally convolved natural image sequences
A. Lőrincz, B. Szatmáry and A. Kabán
Neurocomputing, 38: 1437-1442, 2001
Abstract
It has been reported that independent component analysis (ICA) of natural image sequences yields spatio-temporal filters of non-separable spatio-temporal properties. On the contrary, sign changing filters with separable spatio-temporal properties have not been found via ICA. We show that extending the ICA to temporally convolved inputs develops such receptive fields (RFs). We argue that temporal convolution may arise from the response function of lagged and non-lagged cells of the LGN. The properties of the emerging RFs as a function of convolution time and the dimension of compression are studied.

3. Recognition of novelty made easy: Constraints of channel capacity on generative networks
A. Lőrincz, B. Szatmáry, G. Szirtes, and B. Takács
Connectionist Models of Learning, Development and Evolution, NCPW6 Ed.: R. French (Springer Verlag, London, 2001) pp. 73-82
Abstract
We subscribe to the idea that the brain employs generative networks. In turn, we conclude that channel capacity constraints form the main obstacle for effective information transfer in the brain. Robust and fast information flow processing methods warranting efficient information transfer, e.g. grouping of inputs and information maximization principles need to be applied. For this reason, indepent component analyses on groups of patterns were conducted using (a) model labyrinth, (b) movies on highway traffic and (c) mixed acoustical signals. We found that in all cases 'familiar' inputs give rise to cumulated firing histograms close to exponential distributions, whereas 'novel' information are better described by broad, sometimes truncated Gaussian distributions. It can be shown that upon minimization of mutual information between processing channels, noise can reveal itself locally. Therefore, we conjecture that novelty - as opposed to noise - can be recognized by means of the statistics of neuronal firing in brain areas.

2. Inferior temporal neurons show greater sensitivity to nonaccidental than to metric shape differences
R. Vogels, I. Biederman, M. Bar and A. Lőrincz
Journal of Cognitive Neuroscience 13:444--453, 2001.
Abstract
It has long been known that macaque inferior temporal (IT) neurons tend to fire more strongly to some shapes than to others, and that different IT neurons can show markedly different shape preferences. Beyond the discovery that these preferences can be elicited by features of moderate complexity, no general principle of (nonface) object recognition had emerged by which this enormous variation in selectivity could be understood. Psychophysical, as well as computational work, suggests that one such principle is the difference between viewpoint-invariant, nonaccidental (NAP) and view-dependent, metric shape properties (MPs). We measured the responses of single IT neurons to objects differing in either a NAP (namely, a change in a geon) or an MP of a single part, shown at two orientations in depth. The cells were more sensitive to changes in NAPs than in MPs, even though the image variation (as assessed by wavelet-like measures) produced by the former were smaller than the latter. The magnitude of the response modulation from the rotation itself was, on average, similar to that produced by the NAP differences, although the image changes from the rotation were much greater than that produced by NAP differences. Multidimensional scaling of the neural responses indicated a NAP/MP dimension, independent of an orientation dimension. The present results thus demonstrate that a significant portion of the neural code of IT cells represents differences in NAPs rather than MPs. This code may enable immediate recognition of novel objects at new views.

1. Ockham's razor modeling of the matrisome channels of the basal ganglia thalamocortical loop
A. Lőrincz, Gy. Hévízi and Cs. Szepesvári
International Journal of Neural Systems 11: 125-143, 2001.
Abstract
A functional model of the basal ganglia - thalamocortical (BTC) loops is described. In our modeling effort, we try to minimize the complexity of our starting hypotheses. For that reason, we call this type of modeling Ockham's razor modeling. We have the additional constraint that the starting assumptions should not contradict experimental findings about the brain. First assumption: The brain lacks direct representation of paths but represents directions (called speed fields in control theory). Then control should be concerned with speed-field tracking (SFT). Second assumption: Control signals are delivered upon differencing in competing parallel channels of the BTC loops. This is modeled by extending SFT with differencing that gives rise to the robust Static and Dynamic State (SDS) feedback-controlling scheme. Third assumption: control signals are expressed in terms of a gelatinous medium surrounding the limbs. This is modeled by expressing parameters of motion in parameters of the external space. We show that corollaries of the model fit properties of the BTC loops. The SDS provides proper identification of motion related neuronal groups of the putamen. Local minima arise during the controlling process that works in external space. The model explains the presence of parallel channels as the means to avoiding such local minima. Stability conditions of the SDS predict that the initial phase of learning is mostly concerned with selection of sign for the inverse dynamics. The model provides a scalable controller. State description in external space instead of configurational space reduces the dimensionality problem. Falsifying experiment is suggested. Computer experiments demonstrate the feasibility of the approach. We argue that the resulting scheme has a straightforward connectionist representation exhibiting population coding and Hebbian learning properties.

2000

3. Physiological patterns in the hippocampo-entorhinal cortex system
J.J. Chrobak, A. Lőrincz, and G. Buzsáki
Hippocampus 10: 457-465, 2000
Abstract
The anatomical connectivity and intrinsic properties of entorhinal cortical neurons give rise to ordered patterns of ensemble activity. How entorhinal ensembles form, interact, and accomplish emergent processes such as memory formation is not well-understood. We lack sufficient understanding of how neuronal ensembles in general can function transiently and distinctively from other neuronal ensembles. Ensemble interactions are bound, foremost, by anatomical connectivity and temporal constraints on neuronal discharge. We present an overview of the structure of neuronal interactions within the entorhinal cortex and the rest of the hippocampal formation. We wish to highlight two principle features of entorhinal-hippocampal interactions. First, large numbers of entorhinal neurons are organized into at least two distinct high-frequency population patterns: gamma (40–100 Hz) frequency volleys and ripple (140–200 Hz) frequency volleys. These patterns occur coincident with other well-defined electrophysiological patterns. Gamma frequency volleys are modulated by the theta cycle. Ripple frequency volleys occur on each sharp wave event. Second, these patterns occur dominantly in specific layers of the entorhinal cortex. Theta/gamma frequency volleys are the principle pattern observed in layers I–III, in the neurons that receive cortical inputs and project to the hippocampus. Ripple frequency volleys are the principle population pattern observed in layers V–VI, in the neurons that receive hippocampal output and project primarily to the neocortex. Further, we will highlight how these ensemble patterns organize interactions within distributed forebrain structures and support memory formation.

2. Two-phase computational model of the entorhinal-hippocampal region
A. Lőrincz and Gy. Buzsáki
In: The parahippocampal region: Implications for neurological and psychiatric diseases Eds.: H.E. Sharfman, M.P. Witter, R. Schwarcz. (Annals of the New York Academy of Sciences, Vol. 911, 2000) pp. 83-111.
Abstract
The model described in this chapter is driven by the hypothesis that a major function subserved by the entorhinal cortex (EC)-hippocampal system is to alter the synaptic connections of the neocortex. The model is based on the following postulates. (1) The EC compares the difference between neocortical representations (primary input) and feedback information conveyed by the hippocampus (termed here as the reconstructed input). The difference between the primary input and the reconstructed input (termed 'error') initiates plastic changes in the hippocampal networks. (2) Comparison of the primary input and the reconstructed input requires that these representations are available simultaneously in the EC network. Compensation of time delays can be achieved by means of predictive structures. We suggest that the EC-CA1 connections form such predictive structures. (3) Alteration of synaptic connections in the hippocampus gives rise to new hippocampal outputs, which, in turn, train the long-term memory traces in the EC. We suggest that the goal of long-term memory formation is to simplify structures that represent temporal sequences. We shall argue that neural activities with minimized mutual information (MMI) can meet this challenge. MMI outputs can be generated in a two-step manner, which operations we attribute to the CA3 and CA1 hippocampal fields. (4) The different hippocampal fields can perform both linear and non-linear operations, albeit at different times. The MMI outputs arise under linear operation, whereas long term potentiation of synapses for providing MMI outputs requires non-linear operation. We identify the linear and the non-linear operations with the sharp wave and the theta physiological states of the hippocampus, respectively. We also conjecture that the recurrent collateral system of the CA3 field learns temporal sequences during the theta phase and induces the replay of the learned sequences during the sharp wave phase. (5) We suggest that long-term memory is represented by a distributed and hierarchical reconstruction network, which, owing to the supervised training by the HC, can operate in a single phase and can provide MMI outputs. It then follows, however, that information reaching the hippocampus from these networks may become temporally convolved and that, in turn, would corrupt MMI output formation in the hippocampus itself. We hypothesize that the effect of temporal convolution is counteracted by blind source deconvolution in the excitatory loops of the dentate gyrus. We show that the dentate gyrus can satisfy the strict requirements of the model. memory formation.

1. Modular Reinforcement Learning: A Case Study in a Robot Domain
Zs. Kalmár, Cs. Szepesvári, András Lorincz
Acta Cybernetica 14: 507-522, 2000
Abstract
The behaviour of reinforcement learning (RL) algorithms is best understood in completely observable, finite state- and action-space, discrete-time controlled Markov-chains. Robot-learning domains, on the other hand, are inherently infinite both in time and space, and moreover they are only partially observable. In this article we suggest a systematic design method whose motivation comes from the desire to transform the task-to-be-solved into a finite-state, discrete-time, ''approximately'' Markovian task, which is completely observable, too. The key idea is to break up the problem into subtasks and design controllers for each of the subtasks. Then operating conditions are attached to the controllers (together the controllers and their operating conditions which are called modules) and possible additional features are designed to facilitate observability. A new discrete time-counter is introduced at the ''module-level'' that clicks only when a change in the value of one of the features is observed. The approach was tried out on a real-life robot. Several RL algorithms were compared and it was found that a model-based approach worked best. The learnt switching strategy performed equally well as a handcrafted version. Moreover, the learnt strategy seemed to exploit certain properties of the environment which could not have been seen in advance, which predicted the promising possibility that a learnt controller might overperform a handcrafted switching strategy in the future.

1999

6. Computational model of the entorhinal-hippocampal loop derived from a single principle (Abstract)
A. Lőrincz and Gy. Buzsáki
In: Proceedings of Int. Joint Conf. on Neural Networks, Washington, July 9-16, 1999, IJCNN2136.PDF, IEEE Catalog Number: 99CH36339C, ISBN: 0-7803-5532-6

5. Generative network explains category formation in Alzheimer patients (Abstract)
P. Aszalós, Sz. Kéri, Gy. Kovács, Gy. Benedek, Z. Janka, and A. Lőrincz
In: Proceedings of Int. Joint Conf. on Neural Networks,Washington, July 9-16, 1999, IJCNN2137.PDF, IEEE Catalog Number: 99CH36339C, ISBN: 0-7803-5532-6

4. Winner-take-all network utilizing pseudoinverse reconstruction subnets demonstrates robustness on the handprinted character recognition problem
J. Körmendy-Rácz, Sz. Szabó, J. Lőrincz, Gy. Antal, Gy. Kovács, and A. Lőrincz
Neural Computing and Applications , 8:163-176, 1999.

3. Design and evaluation of a grassfire skeletonization chip (Abstract)
M. Oláh. S. Török, A. Poppe, P. Masa, and A. Lőrincz
In:Proceedings of Conf. MIXDES99, Poland, July, 1999

2. Two-stage computational model of the entorhinal-hippocampal loop (Abstract)
A. Lőrincz, L. Csató, Z. Gábor, and Gy. Buzsáki
Annual Meeting of the Neural Society, Los Angeles, 1999.

1. Parallel and robust skeletonization built from self-organizing elements
Zs. Marczell, Cs. Szepesvári, Zs. Kalmár, and A. Lőrincz
Neural Networks, 12:163--173, 1999.

1998

5. Forming independent components via temporal locking of reconstruction architectures: A functional model of the hippocampus
A. Lőrincz
Biological Cybernetics, 79:263--275, 1998.

4. Basal ganglia perform differencing between `desired' and `experienced' parameters
A. Lőrincz
Computational Neural Science: Trends in Research '97 (Plenum Press) pp. 77-82, 1998.

3. Module-based reinforcement learning: Experiments with a real robot
Zs. Kalmár, Cs. Szepesvári, and A. Lőrincz
Machine Learning 31: 55--85, 1998,
Autonomous Robots 5: 273--295, 1998.

2. Modular reinforcement learning: An application to a real robot task
Zs. Kalmár, Cs. Szepesvári, and A. Lőrincz
Lecture Notes in Artificial Intelligence, 1545: 29-45, 1998.

1. Integrated architecture for motion control and path planning
Cs. Szepesvári, and A. Lőrincz
Journal of Robotic Systems 15: 1--15, 1998.

1997

14. Robust control using inverse dynamics neurocontrollers (Abstract)
Cs. Szepesvári and A. Lőrincz
Nonlinear Analysis Theo., Meth. and Appl. 30: 1669--1676, 1997.

13. The effect of quantum dispersion on laboratory feedback optimal control (Abstract)
G. J. Tóth, A. Lőrincz and H. Rabitz
Journal of Modern Optics 44: 2049--2052, 1997.

12. Abstracting spatial prototypes through short-term suppression of Hebbian weights in a continuously changig environment (Abstract)
S. Tavitian, T. Fomin and A. Lőrincz
Neural Network World 6: 707--727, 1997.

11. Dynamic state feedback neurocontroller for compensatory control (Abstract)
Cs. Szepesvári, Sz. Cimmer and A. Lőrincz
Neural Networks 10: 1691--1708, 1997.

10. Approximate inverse-dynamics based robust control using static and dynamic feedback (Abstract)
Cs. Szepesvári and A. Lőrincz
In: Neural Adaptive Control Theory II . World Scientific, Singapore 1997.

9. Static and dynamic state feedback control model of basal ganglia - thalamocortical loops (Abstract)
A. Lőrincz
International Journal of Neural Systems 8: 339--357, 1997.

8. Common control principles of basal ganglia - thalamocortical loops and the hippocampus (Abstract)
A. Lőrincz
Neural Network World 6: 649--677, 1997.

7. Neurocontrol III: Differencing models of the basal ganglia - thalamocortical loops- (Abstract)
A. Lőrincz
Neural Network World 7:43--72, 1997.

6. Towards a unified model of cortical computation I: Data compression and data reconstruction architecture using dynamic feedback (Abstract)
T. Fomin, J. Körmendy-Rácz and A. Lőrincz
Neural Network World 7:121--136, 1997.

5. Towards a unified model of cortical computation II: From control architecture to a model of consciousness (Abstract)
A. Lőrincz
Neural Network World 7:137--152, 1997.

4. From a control architecture to the basic circuit of the cortex using dynamic state feedback (Abstract)
A. Lőrincz
Sixth Annual Computational Neuroscience Meeting , July 6-10, 1997, Montana, U.S.A. Collection of abstracts p. 96

3. Modelling Huntington's and Parkinson's diseases using dynamic state feedback - (Abstract)
A. Lőrincz
Sixth Annual Computational Neuroscience Meeting , July 6-10, 1997, Montana, U.S.A. Collection of abstracts p. 97

2. A mixed signal VLSI circuit for skeletonization by grassfire transformation (Abstract) (html)
M. Oláh, P. Masa and A. Lőrincz
In Proceedings of ICANN'97, Lausanne, pp. 1205--1210

1. Hippocampal formation trains independent components via forcing input reconstruction (Abstract)
A. Lőrincz
In Proceedings of ICANN'97, Lausanne, pp. 163--168

1996

10. Feasibility of Using Photophoresis to Create a Concentration Gradient of Solvated Molecules (Abstract) Molecular Submarine
B. Space, H. Rabitz, A. Lőrincz and P. Moore.
Journal of Chemical Physics , 105:9515--9524, 1996.

9. Self-Organized Formation of a Set of Scaling Filters and their Neighbouring Connections (Abstract)
T. Rozgonyi, L. Balázs, T. Fomin, and A. Lőrincz
Biological Cybernetics , 75:37--47, 1996.

8. Inverse Dynamics Controllers for Robust Control: Consequences for Neurocontrollers (Abstract)
Cs. Szepesvári and A. Lőrincz
In Proceedings of ICANN'96. Bochum, 1996., pp. 697--702

7. Stabilizing Competitive Learning during On-line Training with an Anti-Hebbian Weight Modulation (Abstract)
S. Tavitian, T. Fomin and A. Lőrincz
In Proceedings of ICANN'96. Bochum, 1996., pp. 791--796

6. Output sensitive discretization for genetic algorithm with migration (Abstract)
Sz. Kovács, G. J. Tóth, R. Der and A. Lőrincz
Neural Network World , 6:101--107, 1996.

5. Approximate Geometry Representations and Sensory Fusion (Abstract)
Cs. Szepesvári and A. Lőrincz
Neurocomputing, 12:267--287, 1996.

4. Generalized skeleton formation for texture segmentation (Abstract)
Zs. Marczell, Zs. Kalmár and A. Lőrincz
Neural Network World, 6:79--87, 1996.

3. Neurocontrol II: High precision control achieved using approximate inverse dynamics models (Abstract)
Cs. Szepesvári and A. Lőrincz
Neural Network World 6:897--920, 1996.

2. Neurocontrol I: Self-organizing speed-field tracking (Abstract)
Cs. Szepesvári and A. Lőrincz
Neural Network World 6:875--896, 1996.

1. Self-organizing multi-resolution grid for motion planning and control (Abstract)
T. Fomin, T. Rozgonyi, Cs. Szepesvári and A. Lőrincz
International Journal of Neural Systems7:757--776, 1996.