Top-down generation of low-resolution representations improves visual perception and imagination.

Area of Science:

• Neuroscience
• Computer Vision
• Artificial Intelligence

Background:

• Visual perception and imagination rely on top-down signals from higher brain areas to the primary visual cortex (V1).
• Top-down signals in V1 possess lower spatial resolution compared to bottom-up visual input.
• The functional advantage of using low-resolution signals for high-resolution representation reconstruction remains unclear.

Purpose of the Study:

• To investigate the role of low-resolution top-down signals in visual perception and imagination.
• To explore the underlying mechanisms by which low-resolution signals facilitate representation reconstruction.
• To apply these findings to improve AI-based sketch generation techniques.

Main Methods:

• Utilized the decoder of a variational auto-encoder (VAE) to model the top-down visual pathway.
• Analyzed the reconstruction of sparse V1 simple cell activities using low-resolution signals.
• Investigated AI-generated sketches, focusing on line thickness and proposing a novel generation technique using blurred sketches with VAEs or GANs.

Main Results:

• Low-resolution top-down signals effectively reconstruct information from sparse V1 simple cell activities, enhancing perception and imagination.
• This low-resolution generation facilitates the formation of geometry-respecting representations in higher cortical areas.
• Sketch generation quality is sensitive to line thickness; thin-line sketches are challenging. A technique using blurred sketches improves thin-line sketch generation.

Conclusions:

• Low-resolution top-down signal processing is a neural strategy for efficient visual perception and imagination.
• This strategy enables the brain to handle sparse representations and form coherent visual percepts.
• Findings offer insights into AI sketch generation, particularly for producing high-quality thin-line images.