AmadeusGPT

AmadeusGPT is an open source natural language interface for analyzing animal behavior, developed in the adaptive intelligence lab led by Mackenzie Mathis. Researchers describe a behavior in plain language, and a large language model such as GPT3.5 or GPT4 writes machine executable code against the AmadeusGPT API, which contains machine learning, computer vision, spatiotemporal reasoning and visualization modules. The aim is behavior analysis without writing code. It is intended for neuroscientists and behavioral researchers who understand their animals but lack deep machine learning or programming skills. To work around the limited context window of language models, it uses a dual memory mechanism in which short term and long term memory communicate through symbols that act as context pointers for retrieval and saving, so distant parts of a conversation are not lost. Users can refine results interactively and add new behavioral modules as they need them. The authors benchmarked it on the MABE 2022 behavior challenge tasks and report leading results. The work, by Shaokai Ye, Jessy Lauer, Mu Zhou, Alexander Mathis and Mackenzie W. Mathis, was published at NeurIPS 2023. AmadeusGPT is free open source software, installed and run on a local machine from its GitHub repository. The project page publishes no pricing, hosted service or commercial plan. It is research software released alongside an academic paper, so it suits labs willing to install and configure the code themselves.