Biologically-inspired neuronal adaptation improves learning in neural networks.

Area of Science:

• Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Human intelligence surpasses artificial neural networks in many tasks.
• Biologically plausible algorithms like Contrastive Hebbian Learning (CHL) and Equilibrium Propagation (EP) offer alternatives to gradient-based methods.
• These algorithms update weights using local information, achieving performance comparable to backpropagation.

Purpose of the Study:

• To augment CHL and EP with a neuronal adaptation mechanism.
• To investigate the impact of this adaptation on artificial neural network performance.
• To explore the potential of neuronal adaptation as a brain mechanism for enhancing learning stability and accuracy.

Main Methods:

• Implemented an 'Adjusted Adaptation' feature inspired by neuronal adaptation.
• Integrated this feature into multilayer perceptrons and convolutional neural networks.
• Trained the modified networks on MNIST and CIFAR-10 datasets.

Main Results:

• The addition of neuronal adaptation surprisingly improved network performance.
• Networks trained with adaptation showed enhanced stability and accuracy.
• The study provides evidence for the benefits of biologically inspired learning mechanisms.

Conclusions:

• Neuronal adaptation is a promising mechanism for improving artificial neural network learning.
• Biologically inspired algorithms can achieve competitive performance with conventional methods.
• Further research into brain mechanisms can lead to more advanced AI.