Complex cells decrease errors for the Müller-Lyer illusion in a model of the visual ventral stream.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Visual perception

Background:

• Artificial visual systems often mimic the human visual cortex for object recognition.
• The Müller-Lyer illusion highlights discrepancies between perception and reality, challenging object recognition systems.
• The Hierarchical Model and Navigation (HMAX) system, a benchmark, exhibits bias with Müller-Lyer stimuli.

Purpose of the Study:

• To investigate how simple and complex cell operations in HMAX influence illusory bias and precision.
• To determine the effect of figure position variation on HMAX's classification of Müller-Lyer images.

Main Methods:

• Two experiments were conducted on the HMAX model.
• Experiment 1 assessed classification bias and uncertainty across HMAX layers, differentiating simple and complex cell functions.
• Experiment 2 evaluated the impact of varying figure positions in input images on HMAX's classification performance.

Main Results:

• Simple cell kernel operations generally increase bias and uncertainty in HMAX.
• Complex cell max-pooling operations tend to decrease bias and uncertainty.
• Increased positional variation in input images reduced bias and uncertainty within HMAX.

Conclusions:

• The Müller-Lyer illusion's effect in HMAX is amplified by simple cells' sensitivity to positional changes.
• Complex cells' robust responses to positional variance help mitigate the illusory bias.
• Findings suggest a link between cell-type specific operations and robustness in artificial visual systems.