Online Cascade Learning for Efficient Inference over Streams

Lunyiu Nie1, Zhimin Ding2, Erdong Hu2, Christopher Jermaine2, Swarat Chaudhuri1,
1The University of Texas at Austin, 2Rice University
Paper arXiv Code Slides

Smaller models in the cascade would continuously improve by imitating the LLM on the harder queries, enabling them to handle an expanding range of queries with increasing proficiency.

Abstract

Large Language Models (LLMs) have a natural role in answering complex queries about data streams, but the high computational cost of LLM inference makes them infeasible in many such tasks. We propose online cascade learning, the first approach to address this challenge.

The objective here is to learn a ``cascade'' of models, starting with lower-capacity models (such as logistic regression) and ending with a powerful LLM, along with a deferral policy that determines the model to be used on a given input. We formulate the task of learning cascades online as an imitation-learning problem, where smaller models are updated over time imitating the collected LLM demonstrations, and give a no-regret algorithm for the problem.

Experimental results across four benchmarks show that our method parallels LLMs in accuracy while cutting down inference costs by as much as 90% with strong robustness against input distribution shifts, underscoring its efficacy and adaptability in stream processing.

BibTeX

@inproceedings{nie2024online,
  title={Online Cascade Learning for Efficient Inference over Streams},
  author={Nie, Lunyiu and Ding, Zhimin and Hu, Erdong and Jermaine, Christopher and Chaudhuri, Swarat},
  booktitle={International Conference on Machine Learning},
  year={2024},
  organization={PMLR}
}