# AutoML-GPT: Automatic Machine Learning with GPT

**Authors:** Shujian Zhang, Chengyue Gong, Lemeng Wu, Xingchao Liu, Mingyuan Zhou  
**arXiv:** [2305.02499](https://arxiv.org/abs/2305.02499)  
**Date:** 2023-05-04  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2305.02499v1/

## Key points

- AutoML-GPT is an Automatic Machine Learning system that uses large language models (LLMs) like GPT-4, LLaMA, Flan-T5, and PaLM.
- LLMs have remarkable capabilities in reasoning, comprehension, and interaction.
- AutoML-GPT automates the process of finding the right model architecture, optimization algorithm, and hyperparameters by leveraging LLMs.
- It acts as a bridge between diverse AI models and dynamically trains models with optimized hyperparameters.
- AutoML-GPT takes user requests and data cards to conduct experiments from data processing to model architecture design, hyperparameter tuning, and predicted training logs.
- The effectiveness of AutoML-GPT is demonstrated in computer vision, natural language processing, and other challenging areas through extensive experiments.
- Building AutoML systems upon GPT significantly improves training efficiency and enhances model performance.
- Use cases across computer vision, natural question answering, and classification benchmarks are showcased.
- AutoML-GPT is effective and general with the potential to create a natural language interface for tuning machine learning models across various tasks.
- Future work includes automatically generating model and data cards for well-known benchmarks as part of the system.

## Summary

The paper presents AutoML-GPT, an Automatic Machine Learning (AutoML) system that utilizes large language models (LLMs) like GPT-4, LLaMA, Flan-T5, and PaLM to automate the training pipeline for various AI tasks. The authors highlight the remarkable capabilities of LLMs in reasoning, comprehension, and interaction. They propose developing task-oriented prompts and leveraging LLMs to automate the process of finding the right model architecture, optimization algorithm, and hyperparameters. AutoML-GPT acts as a bridge between diverse AI models and dynamically trains models with optimized hyperparameters. It takes user requests and data cards to compose prompt paragraphs for conducting experiments from data processing to model architecture design, hyperparameter tuning, and predicted training logs. The authors demonstrate the effectiveness of AutoML-GPT in computer vision, natural language processing, and other challenging areas through extensive experiments and ablation studies. The paper also discusses the benefits of building AutoML systems upon GPT. By automatically conducting machine learning experiments, this approach significantly improves training efficiency and enhances model performance. The authors showcase use cases across computer vision, natural question answering, and classification benchmarks. They further conduct a detailed use case with unseen datasets and additional interactions between users and AutoML-GPT. The proposed AutoML-GPT is deemed effective and general with the potential to create a natural language interface for tuning machine learning models across various tasks. Future work includes automatically generating model and data cards for well-known benchmarks as part of the system and extracting task-aware sub-networks from large pretrained models using ChatGPT. Overall, this paper highlights how LLMs can comprehend natural language effectively to tackle complex AI tasks by automating the training pipeline through AutoML-GPT.