Application:
The rapid growth of neural network models shared on the internet has established model weights as an important data modality. With over 600,000 models on Hugging Face alone and thousands more added daily, there’s an urgent need to organize and understand the relationships between these models. Many are related through a common ancestor (foundation model) from which they were fine-tuned. However, the structure of these relationships remains largely unknown and uncharted, creating a significant gap in our understanding of the AI model ecosystem.
Our Innovation:
We introduce the Model Tree Heritage Recovery (MoTHer) method, a groundbreaking approach to mapping the hereditary relations between AI models.
Our innovation begins with the introduction of Model Tree and Model Graph data structures, which provide a framework for describing the complex relationships between different models. At the heart of our method is a novel algorithm capable of reconstructing Model Trees based solely on model weights, without requiring access to training data or metadata.
Our research has uncovered unique insights into the evolution of weight distributions during model training, which forms the basis of our reconstruction technique. We’ve developed specialized approaches for handling both full fine-tuning and low-rank fine-tuned models, ensuring broad applicability across different training paradigms. This comprehensive approach allows us to uncover the hidden structure of model relationships, providing a powerful tool for understanding the AI model universe.
Advantages:
- High Accuracy: Achieves up to 100% accuracy in reconstructing complex Model Trees
- Versatility: Works with various model types, including vision transformers and language models
- Efficiency: Runs in seconds to minutes, even on CPU
- Real-World Validated: Successfully reconstructed the Llama 2 and Stable Diffusion family trees
- Scalability: Potential for web-scale application across hundreds of thousands of models
Opportunity:
MoTHer Recovery opens up numerous commercial possibilities in the AI industry. It can resolve model ownership disputes, enable a “Google for AI models” search engine, offer model optimization services, ensure compliance in model usage, and enhance AI governance for enterprises. By providing crucial tools for organizing and understanding the expanding universe of AI models, our technology has the potential to become an essential component of the AI development toolkit, driving innovation across the entire ecosystem.
