Abstract:Modular neural networks have shown distinct advantages in many regimes due to the “modular design”. However, how to design an effective decomposition algorithm for improving the prediction accuracy is still a challenge for time series prediction. In this article, a novel modular neural network model based on empirical mode decomposition (EMD) technique and multi-view learning (MNN-EMDMVL) is introduced for time series prediction, which can simultaneously and effectively decompose time series and improve the prediction accuracy of time series. Most of modular neural networks divide time series based on sample space using clustering algorithm without considering its time characteristic, which has an undesirable effect on prediction accuracy. Therefore, the structure of the proposed model consists of two parts: (1) time series decomposition, which can decompose time series into several subseries. (2) network learning, which can further improve the prediction performance of model. Firstly, this paper employs EMD technique to decompose time series into several subseries based on the local characteristic time scale, which can help subnetworks to learn the decomposed subseries more easily. But some subseries with high frequency feature fluctuates greatly, and changes fleetly, which may restrict the performance of subnetworks. Therefore, a novel multi-view learning with winner-take-all strategy (MVWTAS) is proposed to further improve the prediction performance of subnetworks. In MVWTAS, one or more views (features) of subseries with high frequency are selected as the auxiliary tasks and simultaneously learned. Then, the “best model” can be obtained from the auxiliary and main tasks in each time period of the subseries using winner-take-all strategy, thereby improving the prediction accuracy of the subseries with high frequency. Finally, experimental investigations using two benchmark time series and two real-world time series show that compared with other modular neural networks and single modes, MNN-EMDMVL has the distinct advantage in the terms of prediction performance both by time series decomposition using EMD and the developed multi-view learning with winner-take-all strategy.

A Matrix Modular Neural Network Based on Task Decomposition with Subspace Division by Adaptive Affinity Propagation Clustering

CMNN: Coupled Modular Neural Network.

Be expert in multiple aspects and good at many modular neural network with reduced subnet relative training complexity

ModularBoost: an Efficient Network Inference Algorithm Based on Module Decomposition.

Task Decomposition and Modular Single-Hidden-layer Perceptron Classifiers for Multi-Class Learning Problems

Task Decomposition and Module Combination Based on Class Relations: a Modular Neural Network for Pattern Classification

Modular neural network via exploring category hierarchy

MultiModN- Multimodal, Multi-Task, Interpretable Modular Networks

Modularity Facilitates Classification Performance of Spiking Neural Networks for Decoding Cortical Spike Trains

Task Decomposition Based on Class Relations: A Modular Neural Network Architecture for Pattern Classification

A modular neural network with empirical mode decomposition and multi-view learning for time series prediction

A SELF-ORGANIZING NEURAL NETWORK FOR PARTITIONING ON MCM

The Balanced Multi-Modal Spiking Neural Networks with Online Loss Adjustment and Time Alignment

Modularizing while Training: A New Paradigm for Modularizing DNN Models

Multitask Learning With Enhanced Modules.

Modular Boundaries in Recurrent Neural Networks

Neural Network Module Decomposition and Recomposition

Hierarchical Multi-Granularity Attention- Based Hybrid Neural Network for Text Classification.

Co-Evolutionary Modular Neural Networks for Automatic Problem Decomposition

Structured and Balanced Multi-component and Multi-layer Neural Networks

Massively Parallel Classification of Single-Trial EEG Signals Using a Min-Max Modular Neural Network