A single fast Hebbian-like process enabling one-shot class addition in deep neural networks without backbone...

Area of Science:

• Artificial Intelligence
• Neuroscience
• Machine Learning

Background:

• One-shot learning, mimicking a human brain function, is crucial for AI development.
• Current methods like weight imprinting for one-shot learning in deep learning have limitations, including ambiguity in neuroscience relevance and interference with original classifications.
• Advancements in artificial neural networks offer opportunities to model brain functions and develop new AI capabilities.

Purpose of the Study:

• To propose a novel, simple, and practical method for one-shot class addition in deep learning image classifiers.
• To align a weight imprinting process with the Hebbian rule, offering a neuroscience-informed approach.
• To reduce interference with original classifications while enabling learning from single instances.

Main Methods:

• A single Hebbian-like process is employed to enable one-shot class addition in pre-trained deep learning image classifiers.
• Non-parametric normalization is used to mimic the brain's fast Hebbian plasticity.
• The proposed method operates without modifying the original classifier's backbone.

Main Results:

• The Hebbian-like process successfully enables one-shot class addition without altering the classifier's core architecture.
• The use of non-parametric normalization significantly minimizes interference issues previously seen in similar methods.
• The approach demonstrates high practicality and simplicity for one-shot learning tasks.

Conclusions:

• This study presents a simple and effective Hebbian-like mechanism for one-shot class addition in deep learning.
• The method offers a functionally valid hypothesis for neuroscience, linking AI mechanisms to brain function.
• The findings contribute to developing more efficient and interpretable AI systems inspired by biological learning processes.