Neural Network Reprogrammability Tutorial

What You'll Learn

• Understand the unified reprogrammability framework and its theoretical foundations
• Master the taxonomy of input manipulation techniques across different domains
• Apply reprogramming methods to real-world scenarios and evaluate trade-offs
• Assess trustworthiness implications including robustness and security considerations
• Implement practical solutions using provided code examples and frameworks

Why This Tutorial Matters

The massive scale of Foundation Models has created an adaptation bottleneck: how can we efficiently post-train/reuse large-scale pre-trained models for specific tasks without extensive fine-tuning?
Neural Network Reprogrammability answers this question by offering a solution through input space manipulation and output space alignments.

This unified framework encompasses three active research paradigms:

• Model Reprogramming: Learning input transformations to repurpose frozen models
• Prompt Tuning: Learning continuous or discrete prompts to guide model behavior
• Prompt Instruction: Using natural language/visual instructions and few-shot examples

You'll gain both theoretical understanding and practical skills to implement these techniques in your own research and applications.

Tutorial Format

Lecture-based tutorial with interactive demonstrations featuring:

• Theoretical foundations with formal mathematical framework
• Live demonstrations and code walkthroughs
• Case studies across computer vision, NLP, and multimodal domains
• Discussion of trustworthiness implications and best practices
• Hands-on exercises with provided implementation examples