An automated multi parameter neural architecture discovery framework using ChatGPT in the backend.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Designing efficient neural network architectures is complex and requires expertise.
• Edge artificial intelligence (AI) presents additional challenges including power consumption, model size, and inference speed.

Purpose of the Study:

• To introduce a novel framework, LEMONADE, for automated neural network architecture discovery.
• To enable non-AI experts to create efficient models tailored to specific parameters, including edge AI constraints.

Main Methods:

• Developed a framework integrating an expert system and a large language model (LLM) trained on open-domain knowledge.
• Utilized user-defined parameters and considered edge AI constraints without a predefined search space.
• Validated the framework using datasets like CIFAR-10, CIFAR-100, ImageNet16-120, EuroSAT, Malaria Parasite, and IMDb, employing ChatGPT-4o and Gemini-Pro as LLMs.

Main Results:

• Generated neural networks achieved state-of-the-art accuracy on CIFAR-10 and CIFAR-100.
• Demonstrated near state-of-the-art performance on the ImageNet16-120 dataset.
• Successfully created effective neural networks meeting diverse edge AI requirements for datasets like EuroSAT.

Conclusions:

• LEMONADE offers an accessible and effective solution for neural architecture discovery, particularly for edge AI applications.
• The framework's ability to integrate LLMs and expert systems democratizes AI model development.
• LEMONADE shows significant promise in advancing automated machine learning and efficient AI deployment.