Improving large models with small models: Lower costs and better performance.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Pretrained large models (PLMs) offer high performance but incur significant computational costs.
• Reliance on paid APIs for PLMs exacerbates economic burdens for product teams.
• Small models can be effective in specific data distributions or for simpler subtasks.

Purpose of the Study:

• To introduce Data Shunt+ (DS+), a novel paradigm for effective collaboration between small and large AI models.
• To demonstrate how DS+ can reduce the economic burden associated with using PLMs.
• To show that DS+ can enhance the performance of large models by optimizing task allocation.

Main Methods:

• Training small models for various tasks and assessing their confidence levels.
• Developing a query routing mechanism based on small model confidence scores.
• Implementing a collaborative framework where small models handle simpler queries and large models tackle complex ones.

Main Results:

• DS+ significantly reduces the cost of using large models, demonstrated by a 31.18% cost reduction in sentiment analysis.
• DS+ improves accuracy, achieving 95.64% on Amazon Product sentiment analysis compared to ChatGPT's 94.43%.
• The collaborative approach of DS+ is more effective at injecting specific task knowledge into PLMs than traditional fine-tuning.

Conclusions:

• DS+ offers a cost-effective and performance-enhancing solution for leveraging PLMs.
• The paradigm demonstrates the potential of hybrid small-large model systems in AI.
• DS+ provides a superior method for task-specific knowledge integration in PLMs compared to fine-tuning.