Early experience with low-pass filtered images facilitates visual category learning in a neural network model

Area of Science:

• Computational Neuroscience
• Developmental Psychology
• Computer Vision

Background:

• Human infants possess limited contrast sensitivity and visual acuity at birth.
• The functional significance of this early visual immaturity is not fully understood.
• This study investigates potential benefits of reduced visual input during early learning stages.

Purpose of the Study:

• To determine if reduced visual acuity aids in the acquisition of basic-level object categories.
• To examine if this benefit extends to subordinate-level category learning.
• To explore the role of input characteristics (blur, color) in early visual learning.

Main Methods:

• A convolutional neural network (CNN) model was trained using the ecoset dataset.
• Training curricula manipulated blur levels (initial high, gradual reduction) and color (grayscale vs. color).
• Models were subsequently fine-tuned on ImageNet for basic and subordinate category tasks.

Main Results:

• CNNs trained with gradually reduced blur showed improved basic-level categorization compared to those trained without blur.
• Initial blur training enhanced performance in subordinate-level categorization, indicating generalized benefits.
• These advantages were observed only with grayscale inputs, mirroring infant color vision development.

Conclusions:

• Poor visual acuity in newborns may offer functional advantages for visual learning.
• Reduced visual input can facilitate more rapid and accurate acquisition of object categories.
• These findings suggest a potential adaptive role for immature visual systems in early development.