FIR and IIR Synapses, a New Neural Network Architecture for Time Series Modeling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Traditional neural network architectures often face limitations in capturing complex temporal dependencies.
• Feedforward and recurrent structures offer distinct advantages but integrating them effectively remains a challenge.

Purpose of the Study:

• To propose a novel neural network architecture integrating local recurrent and global feedforward components.
• To develop and evaluate a learning rule for this new architecture based on mean square error minimization.
• To compare the performance of the proposed architecture against a standard local-feedforward global-feedforward model.

Main Methods:

• A new neural network architecture was designed, incorporating local recurrent and global feedforward connections.
• A learning rule was derived to minimize the mean square error criterion for training the network.
• Empirical evaluation compared the performance of the local recurrent global feedforward architecture with a local feedforward global feedforward architecture.

Main Results:

• The local recurrent global feedforward neural network architecture was successfully implemented.
• The derived learning rule effectively trained the proposed network.
• The local recurrent global feedforward model demonstrated significantly better performance than the local feedforward global feedforward model.

Conclusions:

• The integration of local recurrent and global feedforward structures offers enhanced capabilities for neural networks.
• The proposed architecture and learning rule provide a promising approach for improving model performance in complex tasks.
• Future research can explore further optimizations and applications of this hybrid architecture.