Reasoning in machine vision by learning fast and slow thinking.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Cognitive Science

Background:

• Human intelligence excels at reasoning for adaptive decision-making in novel situations.
• Current machine intelligence is limited by training data and struggles with dynamic solution refinement during inference.
• Existing machine reasoning advances primarily focus on verbal domains with explicit rules, like mathematical problem-solving.

Purpose of the Study:

• To present a novel paradigm for machine reasoning in computer vision that improves performance with increased inference-time compute.
• To enable performance gains even when labeled training data is scarce.
• To bridge the gap between human adaptive reasoning and current machine intelligence limitations.

Main Methods:

• Developed a dual-process cognitive architecture inspired by human System I (fast) and System II (slow) thinking.
• Integrated a fast-thinking module for initial solution generation and verification.
• Employed a slow-thinking module using self-play reinforcement learning for iterative solution refinement, even with limited task-specific data.

Main Results:

• Demonstrated that extended inference-time compute significantly enhances performance in vision tasks.
• Achieved superior results compared to large-scale supervised learning, foundation models, and human experts.
• Validated the approach on diverse computer-vision benchmarks and a challenging cancer localization task across five organs.

Conclusions:

• Extended inference-time compute offers a powerful mechanism for improving machine reasoning in vision, particularly for data-scarce problems.
• The proposed dual-process paradigm effectively mimics human cognitive strategies for enhanced problem-solving.
• This approach holds significant potential for advancing AI capabilities in complex, real-world visual tasks.