Self-organizing Incremental Neural Networks Research Articles

SOINN+, a Self-Organizing Incremental Neural Network for Unsupervised Learning from Noisy Data Streams

The goal of continuous learning is to acquire and fine-tune knowledge incrementally without erasing already existing knowledge. How to mitigate this erasure, known as catastrophic forgetting, is a grand challenge for machine learning, specifically when systems are trained on evolving data streams. Self-organizing incremental neural networks (SOINN) are a class of neural networks that are designed for continuous learning from non-stationary data. Here, we propose a novel method, SOINN+, which is fundamentally different from the previous generations of SOINN with respect to how “forgetting” is modeled. To achieve a more graceful forgetting, we developed three new concepts: idle time, trustworthiness, and unutility of a node. SOINN+ learns a topology-preserving mapping of the input data to a network structure, which reveals clusters of arbitrary shapes in streams of noisy data. Our experiments with synthetic and real-world data sets showed that SOINN+ can maintain discovered structures even under sudden and recurring concept drifts.

A self-organizing developmental cognitive architecture with interactive reinforcement learning

Developmental cognitive systems can endow robots with the abilities to incrementally learn knowledge and autonomously adapt to complex environments. Conventional cognitive methods often acquire knowledge through passive perception, such as observing and listening. However, this learning way may generate incorrect representations inevitably and cannot correct them online without any feedback. To tackle this problem, we propose a biologically-inspired hierarchical cognitive system called Self-Organizing Developmental Cognitive Architecture with Interactive Reinforcement Learning (SODCA-IRL). The architecture introduces interactive reinforcement learning into hierarchical self-organizing incremental neural networks to simultaneously learn object concepts and fine-tune the learned knowledge by interacting with humans. In order to realize the integration, we equip individual neural networks with a memory model, which is designed as an exponential function controlled by two forgetting factors to simulate the consolidation and forgetting processes of humans. Besides, an interactive reinforcement strategy is designed to provide appropriate rewards and execute mistake correction. The feedback acts on the forgetting factors to reinforce or weaken the memory of neurons. Therefore, correct knowledge is preserved while incorrect representations are forgotten. Experimental results show that the proposed method can make effective use of the feedback from humans to improve the learning effectiveness significantly and reduce the model redundancy.

An Autonomous Developmental Cognitive Architecture Based on Incremental Associative Neural Network With Dynamic Audiovisual Fusion

Developing cognition is difficult to achieve yet crucial for robots. Infants can gradually improve their cognition through parental guidance and self-exploration. However, conventional learning methods for robots often focus on a single modality and train a pre-defined model by large datasets in an offline way. In this paper, we propose a hierarchical autonomous cognitive architecture for robots to learn object concepts online by interacting with humans. Two pathways for audio-visual information are devised. Each pathway has three layers based on the self-organizing incremental neural networks. Visual features and names of objects are incrementally learned and self-organized in an unsupervised way in sample layers, respectively, in which we propose a dynamically adjustable similarity threshold strategy to allow the network itself to control cluster rather than using a pre-defined threshold. Two symbol layers abstract the cluster results from the corresponding sample layer to form concise symbols and transmit them to an associative layer. An associative relationship between two modalities can be built in real time by binding activated visual and auditory symbols simultaneously in the associative layer. In this layer, a top-down response strategy is proposed to let robots autonomously recall another associative modality, solve conflicting associative relationships, and adjust learned knowledge from the top down. The experimental results on two objects datasets and a real task show that our architecture is efficient to learn and associate object view and name in an online way. What is more, the robot can autonomously improve its cognitive level by utilizing its own experience without enquiring with humans.

Time Series Prediction of Tropical Storm Trajectory Using Self-Organizing Incremental Neural Networks and Error Evaluation

This study proposes a route prediction method using a self-organizing incremental neural network. The route trajectory is predicted from two location parameters (the latitude and longitude of the middle of a tropical storm) and the meteorological information (the atmospheric pressure). The method accurately predicted the normalized atmospheric pressure data of East Asia in the topological space of latitude and longitude, with low calculation cost. This paper explains the algorithms for training the self-organizing incremental neural network, the procedure for refining the datasets and the method for predicting the storm trajectory. The effectiveness of the proposed method was confirmed in experiments. With the results of experiments, possibility of prediction model improvement is discussed. Additionally, this paper explains the limitations of proposed method and brief solution to resolve. Although the proposed method was applied only to typhoon phenomena in the present study, it is potentially applicable to a wide range of global problems.

Кластеризация пользователей по голосу с помощью улучшенных самоорганизующихся растущих нейронных сетей

The paper describes the use of an unsupervised learning method based on self-organizing incremental neural networks for the problem of speaker clustering. I uses a set of mel-frequency cepstral coefficients as a user model. This set is obtained by applying a special filter to the sound signal frequency, which was transferred into the mel-frequency scale (mel is an abbreviation of “melody”). The main difference of this work is the consideration of the dynamics of mel-frequency cepstral coefficients changing, which also contains information about the user. The possibility of new unique users emergence in the system while operating makes it impossible to use the major-ity of neural network classes, because learning on a new data set will lead to malfunction, “forgetting” of prior learn-ing. Neural networks for on-line learning impose a limit on the maximum number of clusters, that is unknown for this problem, and, in general, they require a priori knowledge of the input data (to establish thresholds, etc.) that is difficult to achieve in practice. Self-organizing incremental neural networks allow lifetime learning, that means learning during the operation stage, and do not require any a priori knowledge about the users or their quantity. A dynamic neural net-work structure makes it possible to create an unlimited number of new clusters for new previously unregistered users. Thus, this method allows building a flexible speaker clustering system that adapts itself to the changing input data.

Reasoning on the Self-Organizing Incremental Associative Memory for Online Robot Path Planning

SUMMARY Robot path-planning is one of the important issues in robotic navigation. This paper presents a novel robot path-planning approach based on the associative memory using Self-Organizing Incremental Neural Networks (SOINN). By the proposed method, an environment is first autonomously divided into a set of path-fragments by junctions. Each fragment is represented by a sequence of preliminarily generated common patterns (CPs). In an online manner, a robot regards the current path as the associative path-fragments, each connected by junctions. The reasoning technique is additionally proposed for decision making at each junction to speed up the exploration time. Distinct from other methods, our method does not ignore the important information about the regions between junctions (path-fragments). The resultant number of path-fragments is also less than other method. Evaluation is done via Webots physical 3D-simulated and real robot experiments, where only distance sensors are available. Results show that our method can represent the environment effectively; it enables the robot to solve the goal-oriented navigation problem in only one episode, which is actually less than that necessary for most of the Reinforcement Learning (RL) based methods. The running time is proved finite and scales well with the environment. The resultant number of pathfragments matches well to the environment.