One-shot categorization of novel object classes in humans

Area of Science:

• Cognitive Science
• Computer Vision
• Human Perception

Background:

• Human vision excels at generalizing novel object categories from limited examples, a capability not yet matched by artificial intelligence.
• Studying few-shot generalization is difficult as human participants typically see novel samples only once.

Purpose of the Study:

• To investigate human generalization from sparse data using crowdsourcing.
• To compare human generalization abilities with computational models like ShapeComp and AlexNet.

Main Methods:

• Gathered responses from 500 human observers on 20 novel object classes via crowdsourcing.
• Compared human performance to 'ShapeComp' (shape descriptor model) and 'AlexNet' (convolutional neural network).
• Evaluated generalization with 1 or 16 related object samples per novel class.

Main Results:

• Human generalization from sparse data followed systematic patterns related to sample number and variance.
• Both models matched human responses with 16 samples, indicating reliance on shallow, efficient processes.
• With only one sample, models required distinct feature weights, suggesting sophisticated, class-specific processing in humans.

Conclusions:

• Human one-shot object categorization involves active identification of unique class characteristics, surpassing current computational models.
• Sufficient sample data allows observers to use efficient, feature-based generalization processes.