Designing a Deep Learning Framework with Elixir's Axon

Sean Moriarity, the creator of Axon, shares his insights on developing a deep learning framework in Elixir during ElixirConf 2022. Axon distinguishes itself from other frameworks such as TensorFlow and PyTorch by providing an entirely Elixir-based interface, making it simple and familiar to Elixir developers. The API is intentionally small, making it accessible without the need for extensive math or machine learning expertise. Axon is not dependent on NIF or ports but is purely Elixir code, with everything boiling down to Elixir functions. Axon allows for deep learning application development that seamlessly fits within the Elixir environment, whether it is on web servers, IoT devices, or mobile applications. It is designed with portability in mind, enabling easy integration and model imports from TensorFlow and PyTorch. Its simple yet flexible approach allows developers to focus on application development without worrying about underlying complexities, such as maintainability or rapid prototyping. The talk explores key principles behind the design of Axon, its structure, and potential applications, including image classification and language translation, showcasing that deep learning in Elixir is not only possible but can also be effective and efficient.