Anytime Recognition with Routing Convolutional Networks.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Time-sensitive computer vision applications require balancing accuracy and efficiency in deep neural networks.
• Existing 'latest-all' strategies for anytime prediction use fixed exits, which are sub-optimal as earlier exits can be more accurate for certain samples.

Purpose of the Study:

• To improve anytime prediction accuracy by enabling adaptive exit selection for individual samples within a time budget.
• To introduce a novel Routing Convolutional Network (RCN) that dynamically selects the optimal exit layer for each input sample.

Main Methods:

• Developed a Routing Convolutional Network (RCN) that adaptively selects the optimal exit layer for a given time budget.
• Embedded Q-networks at each exit to learn an optimal routing policy, considering information gain and time cost.
• Employed alternate optimization of exits and Q-networks in a cost-sensitive environment to enhance anytime prediction.

Main Results:

• Demonstrated the efficacy of RCN on CIFAR-10, CIFAR-100, and ImageNet classification benchmarks.
• Showcased RCN's adaptability to dense prediction tasks like scene parsing, achieving pixel-level anytime prediction on the Cityscapes benchmark.
• Achieved improved anytime prediction accuracy compared to existing fixed-exit strategies.

Conclusions:

• The proposed RCN effectively improves anytime prediction accuracy by allowing adaptive exit selection per sample.
• RCN offers a flexible and efficient solution for time-sensitive computer vision tasks, applicable to both classification and dense prediction.
• The adaptive routing mechanism provides a superior trade-off between accuracy and efficiency compared to conventional methods.