Learning to Be (In)variant: Combining Prior Knowledge and Experience to Infer Orientation Invariance in Object...

Area of Science:

• Cognitive Science
• Neuroscience
• Computer Vision

Background:

• Object recognition requires visual systems to achieve invariance to transformations like viewpoint changes.
• Orientation invariance is particularly challenging, as it is not universally applicable across all objects.
• Current models often struggle to explain how orientation invariance is learned effectively.

Purpose of the Study:

• To investigate how the visual system learns invariant object recognition, particularly orientation invariance.
• To propose and test a model where invariant transformations are learned by integrating prior expectations with sensory evidence.
• To understand the role of prior expectations in learning orientation dependence.

Main Methods:

• Developed an ideal learner model to simulate the process of learning invariance.
• Designed two behavioral experiments to test the model's predictions.
• Utilized feedback from arithmetic problems to facilitate learning of novel image orientation dependence.

Main Results:

• The ideal learner model predicted that orientation dependence is learned better when prior expectations about orientation are weak.
• Behavioral experiments confirmed this prediction, showing participants learned orientation dependence more effectively under conditions of weak prior expectations.
• The findings support the integration of prior expectations and real-world evidence for learning invariant object representations.

Conclusions:

• The visual system learns invariant object transformations by combining prior expectations with empirical evidence.
• Weak prior expectations facilitate the learning of orientation dependence, suggesting a flexible learning mechanism.
• This framework provides a novel perspective on how the brain achieves robust object recognition despite variations in viewing conditions.