Enhancing Image Model Training with Axon and Nx Libraries

In his presentation for ElixirConf US 2024, Scott Mueller discusses how publicly available Axon/Bumblebee training examples for image models are often less accurate than those produced in other programming languages. He demonstrates new libraries and techniques related to Nx that have the potential to significantly enhance the capabilities of models trained with Elixir. By utilizing the approach showcased in the Elixir Fashion ML Challenge, he walks the audience through essential training techniques and illustrates how to create a custom model using a minimal dataset of around 400 images. After the talk, participants can apply their skills in a Livebook notebook designed for model development, paving the way for more advanced training of language models.