A novel texture backward masking method to locate critical recurrent processes in human vision

Technical Summary: A Novel Texture Backward Masking Method to Locate Critical Recurrent Processes in Human Vision

Problem Statement
Recurrent processes within the visual system are increasingly recognized as critical for performing challenging recognition tasks. Traditionally, backward masking has served as a primary non-invasive tool for investigating these processes; in this paradigm, a mask follows a target image to disrupt ongoing neural computations. However, a significant limitation of conventional backward masking techniques is their inability to pinpoint the specific stage of the visual system where these critical recurrent processes occur. Existing methods disrupt processing broadly but do not allow researchers to isolate whether the disruption happens at early, intermediate, or late stages of the visual hierarchy.

Methodology
To address this spatial ambiguity, the authors propose a novel psychophysics paradigm that leverages recent advances in texture synthesis via deep neural networks. The methodology is grounded in the approximate correspondence between the hierarchical stages of the human visual system and the layers of deep convolutional networks.

The experimental design involves the following components:

1. Target Stimuli: Human subjects are presented with objects under two challenging conditions: either for a very short duration or in unusual poses.
2. Texture Masks: Following the target presentation, subjects are shown textured masks. Crucially, these masks are synthesized to target specific stages of the visual system. The study utilizes masks designed to target only the early visual system and masks designed to target the entire visual system (incorporating features from deeper network layers).
3. Experimental Protocol: The researchers conduct a series of experiments comparing recognition performance when subjects are exposed to these distinct mask types.

Key Contributions
The primary contribution of this work is the development of a new masking technique that moves beyond generic disruption. By synthesizing textures based on deep network layers, the authors create masks that theoretically interact with specific stages of the visual hierarchy. This allows for a more granular investigation of where recurrent processing is recruited during difficult recognition tasks, effectively bridging the gap between deep learning architectures and psychophysical inquiry.

Results
In the conducted experiments, the authors compared the impact of "early-stage" masks against "entire-system" masks on object recognition. The results indicated that both texture types affected recognition abilities to an equal degree. There was no significant difference in performance degradation between the masks targeting only early visual stages and those targeting the entire system.

Significance and Claims
Based on the finding that masks targeting only early stages were as effective as those targeting the entire system, the paper concludes that recurrent processes relevant to these challenging recognition tasks are recruited in, or directed toward, the early stages of the visual system. The authors claim that this novel paradigm successfully identifies the locus of critical recurrent processing, suggesting that for the specific tasks tested (short duration and unusual poses), the necessary recurrent computations occur early in the visual hierarchy. This work provides a methodological framework for localizing recurrent processing without the need for invasive techniques, refining our understanding of the temporal and spatial dynamics of human vision.