Masayasu Atsumi

Saliency-based scene recognition based on growing competitive neural network

This paper proposes the saliency-based scene recognition model in which objects in saliency-based attended spots are sequentially encoded to be invariant with respect to position and size and their positions and sizes are encoded simultaneously. In this model, object recognition and its recall are performed based on the growing two-layered competitive spiking neural network with reciprocal connection between the layers. This neural network represents objects using latency-based temporal coding and grows in size and recognizability through learning and self-organization. Through simulation experiments of a robot equipped with a camera, it is shown that scene recognition is well performed by our model, in which objects are encoded in-variantly with respect to position and size and their positions and sizes are encoded suitably enough for scene recognition.

Simulator Complex for RoboCup Rescue Simulation Project - As Test-Bed for Multi-Agent Organizational Behavior in Emergency Case of Large-Scale Disaster

In the RoboCup Rescue Simulation Project, several kinds of simulator such as Building-Collapse and Road-Blockage Simulator, Fire Spread Simulator and Traffic Flow Simulator are expected to provide a complicated situation in the case of the large-scale disaster through their synergistic effects. It is called Simulator Complex. This article addresses, first, system components of the prototype version of this Simulator Complex, then, explains each of the simulators and the SpaceTime GIS(Geographical Information System) as DBMS(DataBase Management System). In the demonstrations, we have shown the performance enough for a test-bed for multi-agent system development.

Artificial neural development for pulsed neural network design-a simulation experiment on animat's cognitive map genesis

We propose the artificial neural development method that generates the three-dimensional multi-regional pulsed neural network arranged in three layers of the nerve area layer, the nerve sub-area layer, and the cell layer. In this method, the neural development process consists of the first genome-controlled spatiotemporal generation of a neural network structure and the latter spiking activity-dependent regulation of it. For design of genomes, the steady state genetic algorithm is used and it is applied to genomes partially designed manually. Simulation experiments are conducted to generate pulsed neural networks of an animat that moves in an environment. We evolve and develop an animats cognitive map as a multi-regional place recognition circuit that focuses on the place cell area. Through these experiments, we confirm that our method is useful for designing a multi-regional pulsed neural network of the animat that shows biologically realistic features.

A Probabilistic Model of Visual Attention and Perceptual Organization for Constructive Object Recognition

This paper proposes a new probabilistic model of visual attention, figure-ground segmentation and perceptual organization. In this model, spatially parallel preattentive points on a saliency map are organized into sequential selective attended segments through figure-ground segmentation on dynamically-formed Markov random fields and perceptual organization among attended segments are performed in visual working memory for constructive object recognition. Selective attention to segments is controlled based on their saliency, closedness and attention bias. Attended segments in visual working memory are perceptually organized according to a law of proximity. Experiments were conducted by using images of plural categories in an image database and it was shown that selective attention was frequently turned to objects of those categories and that part segments of objects or salient context of objects were perceptually organized.

Attention-Based Segmentation on an Image Pyramid Sequence

This paper proposes a computational model of attention-based segmentation in which a sequence of image pyramids of early visual features is computed for a video sequence and a repetition of selective attention and figure-ground segmentation is performed on the sequence for object perception through successive segment development with mergence of concurrent segments. Attention is stochastically selected on a multi-level saliency map that is called a visual attention pyramid and segmentation is performed on Markov random fields which are dynamically formed around foci of attention. A set of segments and their spatial relation are stored in a visual working memory and maintained through the repetitive attention and segmentation process. Performances of the model are evaluated for basic functions of the vision system such as visual pop-out, figure-ground reversal and perceptual organization and also for real-world scenes which contain objects designed to attract attention.

Probabilistic learning of visual object composition from attended segments

This paper proposes a model of probabilistic learning of object categories in conjunction with early visual processes of attention, segmentation and perceptual organization. This model consists of the following three sub-models: (1) a model of attention-mediated perceptual organization of segments, (2) a model of local feature representation of segments by using a bag of features, and (3) a model of learning object composition of categories based on intra-categorical probabilistic latent component analysis with variable number of classes and intercategorical typicality analysis. Through experiments by using images of plural categories in an image database, it is shown that the model learns a probabilistic structure of intra-categorical composition of objects and context and inter-categorical difference.

Learning visual object categories and their composition based on a probabilistic latent variable model

This paper addresses the problem of statistically learning typical features which characterize object categories and particular features which characterize individual objects in the categories. For this purpose, we propose a probabilistic learning method of object categories and their composition based on a bag of feature representation of cooccurring segments of objects and their context. In this method, multiclass classifiers are learned based on intra-categorical probabilistic latent component analysis with variable number of classes and inter-categorical typicality analysis. Through experiments by using images of plural categories in an image database, it is shown that the method learns probabilistic structures which characterize not only object categories but also object composition of categories, especially typical and non-typical objects and context in each category.

Scene learning and glance recognizability based on competitively growing spiking neural network

We have been building the competitively growing spiking neural network for quick one-shot object learning and glance object recognition, which is the core of our saliency-based scene memory model. This neural network represents objects using latency-based temporal coding and grows size and recognizability through learning and self-organization. Through simulation experiments of a robot equipped with a camera, it is shown that object and scene learning and glance object recognition are well performed by our model.

Sequence learning and planning on associative spiking neural network

We have been building an auto/heteroassociative spiking neural network combined with a working memory model. In this model, a state-driven forward sequence and a goal-driven backward sequence on the associative network are respectively represented by a sequence of synchronous firing in a particular gamma subcycle during a theta oscillation. These forward and backward sequence firings are transmitted to the working memory, temporarily maintained, and integrated based on a competition principle to make a plan. The paper shows that our system can learn forward and backward sequences simultaneously and a plan is incrementally synthesized by repeating their recall and integration.

Scene Memory on Competitively Growing Neural Network Using Temporal Coding: Self-organized Learning and Glance Recognizability

We have been building the competitively growing neural network using temporal coding for quick one-shot object learning and glance object recognition, which is the core of our saliency-based scene memory model. This neural network represents objects using latency-based temporal coding and grows size and recognizability through learning and self-organization. This paper shows that self-organized learning is quickly performed and glance recognition is successfully performed by our model through simulation experiments of a robot equipped with a camera.