You Need Multiple Exiting: Dynamic Early Exiting for Accelerating Unified Vision Language Model

Keyphrases

• Encoder 100%
• Early Exit 100%
• Vision-language Models 100%
• Decoder 33%
• Computational Power 33%
• Unified Architecture 33%
• Inference Efficiency 33%
• Performance Improvement 16%
• Large Animal Model 16%
• Skipping 16%
• Performance Drop 16%
• Confidence Estimation 16%
• Computational Resources 16%
• Model Size 16%
• Confidence-based 16%
• Image Modality 16%
• Input Layer 16%
• MS COCO 16%
• COCO Dataset 16%
• Resource Waste 16%
• Transformer Model 16%
• Text Modality 16%
• Inference Time 16%
• Vision-language Tasks 16%
• Inference Speed 16%

Engineering

• Performance Improvement 100%
• Similarities 100%
• Computational Power 100%
• Intermediate Layer 100%
• Input Layer 100%
• Scale Model 100%

Computer Science

• Language Modeling 100%
• Performance Improvement 50%
• Confidence Estimation 50%
• Computational Power 50%