International congress series最新文献

We fabricated a novel vision system consisting of a neuromorphic silicon retina and FPGA circuits. The silicon retina executes two classes of fundamental image pre-processing: a Laplacian–Gaussian-like spatial filtering and a subtraction of consecutive image frames. The output images of the silicon retina are fed to FPGA circuits, with which image cues, i.e., contours, motion, direction of motion and centre of moving objects, are extracted within a single frame sampling time of the silicon retina. The system has compact hardware and low power consumption and therefore is suitable for controlling autonomous robots.

For the satisfiability problem (SAT), we have proposed a neural network, called LPPH, and a parallelization method, called “mixed parallel execution,” in which LPPH and other algorithms are executed simultaneously. In this paper we study the CPU time of the mixed parallel execution, and prove that the bigger the difference between distribution functions of CPU time of the algorithms used in the parallel execution, the more effective the mixed parallel execution is.

Genetic algorithm (GA) is search algorithm based on the mechanics of natural selection and natural genetics. By the way, hardware accelerators for GA are required to reduce an execution time of GA. In the hardware implementation of GA, a circuit design of roulette wheel selection influences the performance of the GA hardware. In this paper, we propose a new roulette wheel selection circuit based on a rough comparison method (RCM), and evaluate effects of the proposed circuit in terms of the execution time and a circuit size.

In ATR Computational Neuroscience Laboratories, we proposed several computational models such as cerebellar internal models, MOSAIC, and modular and hierarchical reinforcement-learning models. Some of these models can quantitatively reproduce subject behaviors given sensory inputs and reward and action sequences that subjects received and generated. These computational models possess putative information representation such as error signals for internal models and action stimulus dependent reward prediction, and they can be used as explanatory variables in neuroimaging and neurophysiology experiments. We named this approach computational-model-based neuroimaging, as well as computational-model-based neurophysiology. This new approach is very appealing since it is likely the only method with which we can explore neural representations remotely from either sensory or motor interfaces. However, sometimes the limitation of a mere temporal correlation between the theory and data became apparent, so we started to develop a new paradigm, “manipulative neuroscience”, where physical causality is guaranteed.

From an engineering viewpoint, associative memory is one of the most valuable brain functions. A new type of associative memory, morphological associative memory (MAM), has been proposed. The MAM achieves a high perfect recall rate by using a kernel image as an index for pattern recalling. The kernel images, however, are difficult to design for a large number of stored patterns. We developed a block-splitting type morphological associative memory (BMAM) with no need of kernel images. In this paper, the architecture of the BMAM is described and its performance is discussed based on the results of autoassociation experiments.

We investigated the directional sensitivity of four wind-sensitive giant interneurons (GIs), 8-1, 9-1, 9-2 and 9-3, in the fourth-instar nymphs of the cricket Gryllus bimaculatus, using a unidirectional air current stimulus. The directional sensitivity curves expressed with threshold velocities revealed that the preferential directions of GIs in the fourth-instar nymphs were basically the same as those of adults. Intensity-response relations for these GIs in the fourth-, sixth-, last-instar nymphs and adults to air current from front and rear directions were also investigated.

Recent research hypothesizes that the capacity for syntactic recursions forms the computational core of a uniquely human language faculty. Contrary to this hypothesis, Gentner et al. claimed that the capacity to classify sequences from recursive, center-embedded grammar is not uniquely human. We show in this paper that the patterns Gentner used are classified by a Bayesian classifier, a simple and fundamental classifier in machine learning, and consequently we claim that their argument is flawed.

This paper reports on novel developments brought by combining the self-organizing map (SOM) with a modular network. This is called modular network SOM (mnSOM). The mnSOM has an arrayed structure of functional modules aligned on a lattice, and the architecture looks similar to the column structure of our cortex. One of the advantages of the mnSOM is that the user can design the module architecture flexibly depending on the user's purpose, while the backbone algorithm of SOM is kept untouched. This advantage of mnSOM has brought us many variations and applications. In this paper, the concept of mnSOM is first introduced, and variations of mnSOM and their applications are overviewed.

We designed subthreshold analog MOS circuits implementing an inhibitory network model that performs noise-shaping pulse-density modulation (PDM) with noisy neural elements. The aim of our research is to develop a possible ultralow-power delta–sigma type one-bit analog-to-digital converter. Through circuit simulations we confirmed that the signal-to-noise ratio (SNR) of the network was improved by 7.9 dB compared with that of the uncoupled network as a result of noise shaping.

In a recent experimental paper, Foster and Wilson [D.A. Foster, M.A. Wilson, Nature 440 (2006) 680–683] reported reverse replay of recent behavioural sequences in rat hippocampal place cells during the awake state immediately after spatial experience at the reward location. A simple dynamical system firing rate neuron model, illustrating how such replay can generate a map of the environment, selectively linking the replay location with other functionally relevant locations, is presented. Numerical simulations show the rapid development of a cell with receptive field corresponding to the reward location and only one of two cues in a cued T-maze task.