Shape Matters: Few-Shot Object Classification From High-Information Contour Features

Area of Science:

• Computer Vision
• Computational Neuroscience
• Artificial Intelligence

Background:

• Visual object recognition faces challenges in generalizing from limited data and maintaining decision-making transparency.
• Existing models often struggle with robustness and out-of-distribution generalization.

Purpose of the Study:

• To develop a biologically inspired model for visual object recognition that generalizes from limited data and offers transparent decision-making.
• To enhance model robustness and generalization capabilities, mimicking human concept learning.

Main Methods:

• Proposed a model classifying images based on transformation-invariant local shape key features.
• Encoded features using image patches and polar coordinates for interpretable comparisons.
• Utilized clustering for prototype selection to improve representativeness and mimic human concept learning.

Main Results:

• Achieved human-comparable performance with 1-2% error rate on MNIST using all training images as prototypes.
• Outperformed convolutional neural networks (CNNs) in data-limited scenarios with few prototypes.
• Demonstrated strong out-of-distribution generalization on the ETL-1 dataset, maintaining high accuracy where CNNs faltered.

Conclusions:

• The biologically inspired model offers robust and generalizable visual object recognition, outperforming CNNs in challenging scenarios.
• The model's approach, using local shape features and prototype selection, brings AI closer to human-like recognition capabilities.
• This work highlights the potential of bio-inspired computing for advancing AI in visual recognition tasks.