Search

Abstract Deep learning neural networks produce excellent results in various pattern recognition tasks. It is of great practical importance to answer some open questions regarding model design and parameterization, and to understand how input data are converted into meaningful knowledge at the output. The layer-by-layer evolution of the abstraction level has been proposed previously as a quantitative measure to describe the emergence of knowledge in the network. In this work we systematically evaluate the abstraction level for a variety of image datasets. We observe that there is a general tendency of increasing abstraction from input to output with the exception of a drop of abstraction at some ReLu and Pooling layers. The abstraction level is relatively low and does not change significantly in the first few layers following the input, while it fluctuates around some high saturation value at the layers preceding the output. Finally, the layer-by-layer change in abstraction is not normally distributed, rather it approximates an exponential distribution. These results point to salient local features of deep layers impacting overall (global) classification performance. We compare the results extracted from deep learning neural networks performing image processing tasks with the results obtained by analyzing brain imaging data. Our conclusions may be helpful in future designs of more efficient, compact deep learning neural networks. [ABSTRACT FROM AUTHOR]

In recent decades a significant effort has been directed towards experimental and theoretical approaches aiming at the establishment of neural correlates of higher cognitive functions and awareness. These efforts produce massive amounts of data on imaging brain structure and functions, and efficient methods are in high demand to make these data easily accessible and understandable to human experts. Here we describe the development of qualitative tools and methodologies where large quantities of brain imaging data are processed and displayed for the purpose of visually discriminating the various stages of the cognitive processes. In this work we report and describe in detail a methodology inspired by the art of encephalography, whereby brain dynamics movies are created based on experimental data. The results are presented to identify large-scale synchronizations and de- synchronizations across broad frequency bands as a potential manifestation of the cycle of creation of knowledge and meaning. This visual process is also useful for the description of learning and adaptation processes in brains. [ABSTRACT FROM AUTHOR]

This paper provides new stability results for Action-Dependent Heuristic Dynamic Programming (ADHDP), using a control algorithm that iteratively improves an internal model of the external world in the autonomous system based on its continuous interaction with the environment. We extend previous results for ADHDP control to the case of general multi-layer neural networks with deep learning across all layers. In particular, we show that the introduced control approach is uniformly ultimately bounded (UUB) under specific conditions on the learning rates, without explicit constraints on the temporal discount factor. We demonstrate the benefit of our results to the control of linear and nonlinear systems, including the cart–pole balancing problem. Our results show significantly improved learning and control performance as compared to the state-of-the-art. [ABSTRACT FROM AUTHOR]

Mathematical approaches are reviewed to interpret intermittent singular space–time dynamics observed in brain imaging experiments. The following aspects of brain dynamics are considered: nonlinear dynamics (chaos), phase transitions, and criticality. Probabilistic cellular automata and random graph models are described, which develop equations for the probability distributions of macroscopic state variables as an alternative to differential equations. The introduced modular neuropercolation model is motivated by the multilayer structure and dynamical properties of the cortex, and it describes critical brain oscillations, including background activity, narrow-band oscillations in excitatory–inhibitory populations, and broadband oscillations in the cortex. Input-induced and spontaneous transitions between states with large-scale synchrony and without synchrony exhibit brief episodes with long-range spatial correlations as observed in experiments. [ABSTRACT FROM AUTHOR]

Abstract: Sensory information processing and cognition in brains are modeled using dynamic systems theory. The brain’s dynamic state is described by a trajectory evolving in a high-dimensional state space. We introduce a hierarchy of random cellular automata as the mathematical tools to describe the spatio-temporal dynamics of the cortex. The corresponding brain model is called neuropercolation which has distinct advantages compared to traditional models using differential equations, especially in describing spatio-temporal discontinuities in the form of phase transitions. Phase transitions demarcate singularities in brain operations at critical conditions, which are viewed as hallmarks of higher cognition and awareness experience. The introduced Monte-Carlo simulations obtained by parallel computing point to the importance of computer implementations using very large-scale integration (VLSI) and analog platforms. [Copyright &y& Elsevier]

Abstract: Large-scale networks with hundreds of thousands of variables and constraints are becoming more and more common in logistics, communications, and distribution domains. Traditionally, the utility functions defined on such networks are optimized using some variation of Linear Programming, such as Mixed Integer Programming (MIP). Despite enormous progress both in hardware (multiprocessor systems and specialized processors) and software (Gurobi) we are reaching the limits of what these tools can handle in real time. Modern logistic problems, for example, call for expanding the problem both vertically (from one day up to several days) and horizontally (combining separate solution stages into an integrated model). The complexity of such integrated models calls for alternative methods of solution, such as Approximate Dynamic Programming (ADP), which provide a further increase in the performance necessary for the daily operation. In this paper, we present the theoretical basis and related experiments for solving the multistage decision problems based on the results obtained for shorter periods, as building blocks for the models and the solution, via Critic-Model-Action cycles, where various types of neural networks are combined with traditional MIP models in a unified optimization system. In this system architecture, fast and simple feed-forward networks are trained to reasonably initialize more complicated recurrent networks, which serve as approximators of the value function (Critic). The combination of interrelated neural networks and optimization modules allows for multiple queries for the same system, providing flexibility and optimizing performance for large-scale real-life problems. A MATLAB implementation of our solution procedure for a realistic set of data and constraints shows promising results, compared to the iterative MIP approach. [Copyright &y& Elsevier]

Human cognition performs granulation of the seemingly homogeneous temporal sequences of perceptual experiences into meaningful and comprehensible chunks of fuzzy concepts and behaviors. These knowledge granules are stored and consequently accessed during action selection and decisions. A dynamical approach is presented here to interpret experimental findings using K (Katchalsky) models. In the K model, meaningful knowledge is repetitiously created and processed in the form of sequences of oscillatory patterns of neural activity distributed across space and time. These patterns are not rigid but flexible and intermittent; soon after they arise through phase transitions, they dissipate. Computational implementations demonstrate the operation of the model based on the principles of intentional brain dynamics. [Copyright &y& Elsevier]

Abstract: We present an intentional neurodynamic theory for higher cognition and intelligence. This theory provides a unifying framework for integrating symbolic and subsymbolic methods as complementary aspects of human intelligence. Top-down symbolic approaches benefit from the vast experience with logical reasoning and with high-level knowledge processing in humans. Connectionist methods use bottom-up approach to generate intelligent behavior by mimicking subsymbolic aspects of the operation of brains and nervous systems. Neurophysiological correlates of intentionality and cognition include sequences of oscillatory patterns of mesoscopic neural activity. Oscillatory patterns are viewed as intermittent representations of generalized symbol systems, with which brains compute. These dynamical symbols are not rigid but flexible. They disappear soon after they emerged through spatio-temporal phase transitions. Intentional neurodynamics provides a solution to the notoriously difficult symbol grounding problem. Some examples of implementations of the corresponding dynamic principles are described in this review. [Copyright &y& Elsevier]

Abstract: The hierarchy of models of interacting neural populations with excitatory and inhibitory connections is described using second-order nonlinear ordinary differential equations. Systematic analytical and numerical studies are aimed at determining stability conditions of the coupled system. Generalized stability conditions are derived for a class of excitatory–inhibitory neural population models. Equilibrium continuation analysis is applied to interpret the obtained stability conditions. The results are discussed in the context of chaotic brain dynamics theory and chaotic itinerancy. Attractor switching in biologically relevant multi-stable systems is demonstrated. [Copyright &y& Elsevier]

In response to this digital divide between developed and developing countries in their use of computers to prepare students for the global economy, the World Bank and, subsequently, the World Links organization provided schools in developing countries with networked computers and training that supports integration of ICT into teaching. This article synthesizes findings from 3 years of evaluative research on the program. The findings are based on surveys of teachers, students, administrators, and technology coordinators, as well as a field test assessment of student learning. The study examines the services provided, documents their impact, and draws implications for policy in developing countries. [Copyright &y& Elsevier]

A dynamical classification method is introduced based on the principle of encoding sensory information in oscillating spatio-temporal patterns. The method is used for the evaluation of EEG signals measured by spatially distributed electrodes over sensory cortices of rabbits. Our analysis reveals two types of patterns in neocortex. One type occurs with short latency after stimulus arrival. This early pattern represents the direct impact of a discriminated stimulus on the receiving cortex. The other type is endogenous and occurs with variable latency in the subsequent ∼1200 ms. It represents the destabilization of the spatio-temporal dynamics of the cortices and consequent phase transitions. [Copyright &y& Elsevier]

Abstract: The present Special Issue “Advances in Neural Networks Research: IJCNN2009” provides a state-of-art overview of the field of neural networks. It includes 39 papers from selected areas of the 2009 International Joint Conference on Neural Networks (IJCNN2009). IJCNN2009 took place on June 14-19, 2009 in Atlanta, Georgia, USA, and it represents an exemplary collaboration between the International Neural Networks Society and the IEEE Computational Intelligence Society. Topics in this issue include neuroscience and cognitive science, computational intelligence and machine learning, hybrid techniques, nonlinear dynamics and chaos, various soft computing technologies, intelligent signal processing and pattern recognition, bioinformatics and biomedicine, and engineering applications. [Copyright &y& Elsevier]

Deep Reinforcement Learning (RL) demonstrates excellent performance on tasks that can be solved by trained policy. It plays a dominant role among cutting-edge machine learning approaches using multi-layer Neural networks (NNs). At the same time, Deep RL suffers from high sensitivity to noisy, incomplete, and misleading input data. Following biological intuition, we involve Spiking Neural Networks (SNNs) to address some deficiencies of deep RL solutions. Previous studies in image classification domain demonstrated that standard NNs (with ReLU nonlinearity) trained using supervised learning can be converted to SNNs with negligible deterioration in performance. In this paper, we extend those conversion results to the domain of Q-Learning NNs trained using RL. We provide a proof of principle of the conversion of standard NN to SNN. In addition, we show that the SNN has improved robustness to occlusion in the input image. Finally, we introduce results with converting full-scale Deep Q-network to SNN, paving the way for future research to robust Deep RL applications. [ABSTRACT FROM AUTHOR]

In recent years, spiking neural networks (SNNs) have demonstrated great success in completing various machine learning tasks. We introduce a method for learning image features with locally connected layers in SNNs using a spike-timing-dependent plasticity (STDP) rule. In our approach, sub-networks compete via inhibitory interactions to learn features from different locations of the input space. These locally-connected SNNs (LC-SNNs) manifest key topological features of the spatial interaction of biological neurons. We explore a biologically inspired n -gram classification approach allowing parallel processing over various patches of the image space. We report the classification accuracy of simple two-layer LC-SNNs on two image datasets, which respectively match state-of-art performance and are the first results to date. LC-SNNs have the advantage of fast convergence to a dataset representation, and they require fewer learnable parameters than other SNN approaches with unsupervised learning. Robustness tests demonstrate that LC-SNNs exhibit graceful degradation of performance despite the random deletion of large numbers of synapses and neurons. Our results have been obtained using the BindsNET library, which allows efficient machine learning implementations of spiking neural networks. [ABSTRACT FROM AUTHOR]