Researchers have discovered that denoising models, when trained on natural images, develop internal representations that are sensitive to perceptual illusions, similar to human observers. These representations were found in specific internal layers across various model architectures, with the denoising objective being a more significant driver than the architecture itself. While these internal representations correlate with psychophysical models of human perception and can be causally linked to internal signal processing through channel ablation, they do not affect the model's output, leading the researchers to term them "perceptual phantoms." AI
IMPACT Reveals that internal model representations can mimic human perception, even when not reflected in output, suggesting new avenues for understanding and evaluating AI.
RANK_REASON Academic paper detailing novel findings in AI model representations. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
