How well do rudimentary plasticity rules predict adult visual object learning?

Area of Science:

• Cognitive Neuroscience
• Computer Vision
• Machine Learning

Background:

• A fundamental challenge in visual object learning is identifying novel objects from a limited set of training images.
• A prominent hypothesis suggests the adult brain solves this via neural re-representation and optimized decision boundaries using plasticity rules.
• Previous research lacked image-computable models and naturalistic human learning data to test this hypothesis.

Purpose of the Study:

• To bridge the gap between computational models and human behavior in visual object learning.
• To develop and release empirical benchmarks for evaluating object learning models against human performance.
• To investigate the explanatory power of different visual representations and plasticity rules in human learning.

Main Methods:

• Developed 2,408 testable learning models based on contemporary image-computable models of the primate ventral visual stream.
• Collected human learning trajectories through online psychophysics on tasks involving novel 3D objects (371,000 trials).
• Evaluated learning models against human behavioral benchmarks to assess their alignment with human learning.

Main Results:

• Models utilizing deep, high-level representations derived from neural networks showed significant alignment with human learning behavior.
• No single model fully explained all replicable human behavior, indicating the complexity of the learning process.
• Rudimentary plasticity rules demonstrated high explanatory power when paired with appropriate visual representations.

Conclusions:

• Deep, high-level visual representations are crucial for explaining human object learning from limited data.
• Simple plasticity rules, when applied to suitable representations, can effectively predict human learning trajectories.
• The study provides a framework and benchmarks for future research in computational models of visual cognition.