This study analyzes the role of Chaotic Neural Networks (ChNN) in EVSF-Network, a feature binding learning model, from an information-theoretic perspective. EVSF-Network is a hybrid architecture combining Convolutional Neural Networks (CoNN), ChNN, and Recurrent Neural Networks (RNN), implementing feature binding in images through synchronization states in ChNN. Based on Information Bottleneck theory, we analyze ChNN’s role from the perspectives of information compression and reconstruction. Specifically, we established and experimentally verified three hypotheses: (1) weak chaotic states in ChNN achieve optimal feature binding, (2) unit group formation based on synchronous oscillations enables efficient information compression, and (3) temporal evolution plays an essential role in EVSF-Network learning through reconstruction. Experimental results demonstrate that optimal learning occurs when ChNN is in a weak chaotic state, where information compression and reconstruction are appropriately balanced. Analysis of temporal evolution reveals that moderate embedding dimensions in weak chaotic states enable effective information integration. These findings indicate that ChNN plays a crucial role in EVSF-Network learning through both information compression and reconstruction.

EVSF-Network is a deep neural network for learning dynamic symbols. This network was a hybrid neural network configured from a convolutional network and a chaotic neural network. The convolutional network learns figures and their features from image data, while the chaotic neural network detects feature binding information of figures. However, the network has a problem where the learning performance deteriorates when the number of figures in the image is indefinite. To address this problem, a sub-network using a recurrent network has been added. This recurrent network uses the detected binding information as attention and reflects it in the output. The procedure for detecting feature bindings is described first. Information about feature bindings is detected from the internal state of the output of the intermediate layer of the convolutional network, which serves as input to the chaotic neural network. Here, the property of the chaotic neural network as a coupled chaotic lattice is utilized. The feature bindings are detected as synchronous oscillations between chaotic units. Next, a method is described to reflect the detected binding information in the output as attention. An overview of EVSF-Network and its learning procedure is presented, followed by experiments and results. The experiments are based on the image data used by Rosenblatt when he posed the binding problem. Finally, a discussion based on the experimental results is presented, and future work is outlined, focusing on how to deal with the symbol grounding problem.

EVSF-Network is a deep neural network that processes dynamic symbols. The dynamic symbols change their meaning depending on the situation. The network is a hybrid neural network that combines a convolutional network and a chaotic neural network. The chaotic neural network detects binding information between features from the state of all binding layers of the convolutional network. The detected binding information is reflected in the output of the convolutional network as an attention. This paper presents the core features of the EVSF-Network, namely the feature co-occurrence detection and the learning. First, the properties of dynamic symbols are presented through a discussion of the binding problem. Next, the structure of EVSF-Network and how it is learned are presented. There, a scheme is presented in which synchronization groups (unit clusters) detected from the states of fully connected layer of a convolutional neural network converted to pseudo-time series are treated as binding information. Following that, an experiment with a task based on the binding problem and its results are presented. Finally, based on the experimental results, a discussion of some issues and future challenges are presented.

In this paper, EVSF-Network, a deep neural network for dynamic symbols, is presented. EVSF-Network is an extension of VSF-Network, which was developed for the same purpose. The symbols that can be processed by VSF-Network are static symbols, while those that can be processed by EVSF-Network are dynamic symbols. Unlike static symbols, dynamic symbols have the property of changing their meaning according to the situation. It is a hybrid neural network that combines a hierarchical neural network and a chaotic neural network. The hierarchical neural networks perform feature extraction. The chaotic neural network extracts co-occurrence relations between perceptual features depending on situations from the internal state of the hierarchical neural network. The co-occurrence relations between features found by the chaotic neural network are reflected in the output of the hierarchical neural network by attention. First, the binding problem is discussed as a problem related to basic properties about dynamic symbols. A simple example of the binding problem is introduced, followed by an explanation of the synchronization group of neurons, a neuroscience model of the binding problem. EVSF-Network expresses situation dependency in dynamic symbols by means of synchronization groups. Next, the structure of EVSF -Network and its learning procedure are presented. Following that, an experiment using an image recognition task dealing with the binding problem by Rosenblatt et al. and their results are presented. Finally, a discussion about future extensions to the EVSF-Network is shown.

In this paper, a theoretical analysis on the internal structure and learning method of VSF-Network is introduced. It is a hybrid neural network combining hierarchical neural network and the chaotic neural network. The hierarchical neural network learns patterns and recognizes patterns based on the learning results.. The chaotic neural network monitors the internal structure of the hierarchical neural network and identifies learned units and unused units. The result of this identification is used for selection of the updating target at the time of updating the weight. With this selective weight update, the hierarchical neural network of the VSF-Network is divided into subnetworks that recognize previously learned patterns and subnetworks that recognize newly learned patterns. The monitoring by chaotic neural network is explained as calculation of eigenspace of its recall process. Furthermore, it is explained from the viewpoint of the combination of the divided linear spaces that the subnetworks obtained by learning can be recognized in combination according to the situation. Finally, the validity and problems of the theoretical explanation are introduced through the analysis of the intermediate layer of VSF-Network during incremental learning.

VSF-Network is a neural network model that learns dynamical patterns. It is hybrid neural network combining a chaotic neural network and a hierarchical neural network. The hierarchical neural network part is used for pattern learning. The chaotic neural network part monitors behavior of neurons in the middle layer of the hierarchical neural network. In this paper, two theoretical backgrounds of VSF-Network are introduced. An incremental learning model using chaotic neural networks is introduced. The monitoring by chaotic neural network is based on the clusters of synchronized oscillators. Using the monitoring results, redundant neurons in the hierarchical neural network are found and they are used for learning of new patters. The second background is about the pattern recognition by combining learned patterns. The mechanism about recognition of combined learned patterns is explained by subspace selection in linear space. Through an experiment, its ability for the incremental learning and the pattern recognition are shown, and the factors influencing learning of VSF-Network are also shown.

VSF-Network, Vibrate Synchronize Function Network, is a hybrid neural network combining a Chaos Neural Network with a hierarchical network. VSF-Network is designed for symbol learning by a neural network. It finds unknown parts of input data by comparing to learned pattern and it learns unknown patterns using unused part of the network. The new patterns are learned incrementally and they are represented as sub-networks with unused parts of hierarchical neural network. Combinations of patterns are represented as combinations of the sub-networks. The combinations of symbols are represented as combinations of the sub-networks. In this paper, the two theoretical backgrounds of VSF-Network are introduced. At the first, an incremental learning framework with Chaos Neural Networks is introduced. Next, the pattern recognition with the combined with symbols is introduced. By Stochastic Catastrophe Model, the authors explain the combined pattern recognition. Through an experiment, both the incremental learning capability and the pattern recognition with pattern combination. Index Terms: Incremental learning, Chaos Neural network, Nonlinear dynamics, Stochastic Catastrophe Model.

In this paper, a neural network model, which learns symbols is introduced. VSF-Network (Vibration Synchronizing Function Network) is a hybrid neural network combining a chaos neural network with a hierarchical network. It has an ability for a incremental learning of patters by abstracting input data. VSF-Network finds unknown elements in data based on clusters generated by chaos neurons. VSF-Network generates sub-networks while it learns new patterns based on the information about the clusters. When a combination of learned patterns is presented, VSF-network recognizes them by combining its sub-networks. In this paper, an incremental learning model to examine the dynamics of VSF-network is introduced. The performances of the incremental learning by VSF-network are shown through the two tasks and the discussion about the combination form of the sub-networks generated by VSF-Network.

In this paper, we show an overview of VSF-network, the presumption of parameters for the additive learning, results of the learning applied to obstacle avoidance task using the presumed parameters, and we examined the state of the hidden-layer in VSF-network that the additive learning is applied. The recognition of patterns that are the learned the existing pattern, the incrementally learned pattern, and the pattern that is combined those both patterns, are improved, by setting the state of GCM-module where is a weak chaotic state in the incremental learning phase. The feature which can be recognized using the pattern that combines both the freshly learned pattern and the existing pattern that have never learned, is the key feature of VSF-network. A T-junction, a simple obstacle, and a compound obstacle were provided to a hierarchical network and VSF network that are incrementally learned, and the outputs from the hidden-layer were compared. Through the comparison, we confirmed that the output pattern of units that is incrementally learned pattern, and the combination of both patterns respectively on VSF-network.

In this paper, dynamics of VSF-network (vibration synchronizing function network) is investigated. VSF-network is a model of neural networks that segments information from external world and fixes the segmented information. VSF-network is a hybrid neural network, and the chaos neuron is used for the hidden layer of it. VSF-network articulates information with the neuron cluster generated by the synchronizing vibration that the chaos neuron in hidden layer shows. We analyze the dynamics of generating the neuron cluster. Factors affecting cluster generation are investigated. The stability and the bifurcation of the neuron cluster and factors affecting cluster generation are investigated.

This paper proposes a new demand forecasting method using the neural network and Fourier transform. In this method, time series data of sales results considered as a combination of frequency are transformed into several frequency data. They are identified from objective indexes that consist of product properties or economic indicators and so forth. This method is efficient for demand forecasting aimed at new products that have no historical data.

When a system tries to recognize its external world, it should segment information on the 'world'. In this paper, we propose the vibration synchronize function network (VSF-network) that is a neural network model for segmenting information from the external world. We start with a discussion of the relation of information encoded into neural networks and its situation. In the next paragraph an overview of VSF-network and learning algorithm of VSF-network are given. Our VSF-network has been applied for the behavior acquisition of a rover avoiding obstacles. Finally, we discuss performances of VSF-network observed in this application.

This paper proposes a new demand forecasting method using a neural network and Fourier transform. In this method, time series data of sales results considered as a combination of frequency are transformed into several frequency data. They are identified from objective indexes that consist of product properties or economic indicators and so forth. This method is efficient for demand forecasting aimed at new products that have no historical data.