A 3D ray traced biological neural network learning model.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Training large neural networks demands substantial computational resources and time.
• Current transfer learning methods are constrained by fixed model dimensions, limiting architectural flexibility for new datasets.

Purpose of the Study:

• To develop a dynamic neural network inspired by biological neural networks (BNNs) that overcomes the limitations of current transfer learning algorithms.
• To create a transferable learning model adaptable to various network architectures and datasets.

Main Methods:

• Designed a dynamic neural network employing raytracing to connect neurons in 3D space, enabling flexible network growth.
• The network architecture can dynamically adjust its shape and size to accommodate different datasets and environments.

Main Results:

• The proposed transfer learning algorithm demonstrated the fastest training times on the Alcala dataset across varying environments and input sizes.
• Achieved superior performance compared to state-of-the-art methods on an EEG dataset.

Conclusions:

• The dynamic neural network offers a flexible and efficient approach to transfer learning, adaptable to diverse architectures and datasets.
• This method shows promise for future implementation in real biological neural networks to reduce power consumption.