Hi, I’m writing to submit an abstract for a proposed talk for Apache MXNet day.
Thanks for your consideration! Jonas Talk Title: AutoGluon: Powerful and Easy-to-use AutoML for Text, Image, and Tabular Data Talk Abstract: AutoGluon<https://github.com/awslabs/autogluon/> is an open-source software toolkit to automate machine learning (ML) on image, text, and tabular data (built on top of MXNet). With a couple lines of Python code, anybody can translate their raw data into highly accurate predictions, regardless of their ML expertise. In comprehensive benchmarks<https://arxiv.org/pdf/2003.06505.pdf> over 50 datasets, AutoGluon produced significantly more accurate models (given the same runtime/compute) than other popular AutoML tools like TPOT, H2O, AutoWEKA, auto-sklearn, Google AutoML Tables. In two prominent Kaggle competitions with tabular data, AutoGluon outperformed 99% of human data science teams after just being run for 4 hours on the raw data. In comparisons with human data scientists on 4 image classification competitions from Kaggle<https://medium.com/@zhanghang0704/image-classification-on-kaggle-using-autogluon-fc896e74d7e8>, AutoGluon consistently ranked around the top 10%. AutoGluon has also been used to win a number of internal prediction competitions at Amazon, and is being used in production for multiple applications. This talk describes how AutoGluon achieves such strong predictive performance through state-of-the-art deep learning, hyperparameter-optimization, stack ensembling, and other modelling techniques.
