Abstract

Model optimization plays the key role in many learning and vision tasks. However, designing numerical schemes always need high mathematical skills and rich domain knowledge. Moreover, how to apply the generally designed iterations in specific real-world scenario is always a challenging problem. In this talk, we introduce a series of paradigms to design task-specific optimization schemes based on (inexact) learnable architectures. The theoretical properties of these deeply trained propagations are carefully investigated. We demonstrate that we actually provide a new way to establish interpretable and collaborative deep learning models for different real-world applications. Some insights (e.g., the comparison to adversarial mechanism in GAN) will also be covered.