Constrained sampling from deep generative image models reveals mechanisms of human target detection

Area of Science:

• Computational neuroscience
• Computer vision
• Human visual perception

Background:

• Early visual processing models, like oriented filters and divisive normalization, excel in simple tasks.
• Their efficacy in complex, naturalistic image processing remains less understood.

Purpose of the Study:

• To evaluate how well existing computational models of early vision predict human performance in localizing targets within complex natural images.
• To compare biologically inspired models against deep convolutional neural networks (CNNs) for this task.

Main Methods:

• Embedding arc segments of varying curvature into naturalistic images using a deep generative model.
• Human observers localized these arc targets, with data analyzed against multiple computational models.
• Models included oriented filter banks, CNNs, and biologically inspired variants with normalization or gain control.

Main Results:

• Four models showed strong predictive power: oriented filters, filters with tuned gain control, filters with local normalization, and a trained CNN.
• A control experiment revealed that the oriented filters with tuned gain control model best explained human observer data.
• This suggests limitations in applying standard CNNs to complex visual tasks compared to biologically constrained models.

Conclusions:

• Standard models of early vision, particularly those incorporating cortical surround interactions, generalize well to complex visual tasks.
• Biologically inspired models offer superior predictions of human performance in naturalistic image perception compared to general-purpose CNNs.