Elixir Mentor Sean Moriarity on Machine Learning in Elixir
Jun 16, 2024
Sean Moriarity, author and Elixir machine learning practitioner known for books on genetic algorithms and Nx/Axon work. He talks about using genetic algorithms for optimization, why Elixir fits ML workflows, and the current Nx/Axon stack. Short takes cover data representation, Livebook advantages, adapting pre-trained models, deployment trade-offs, and gaps like quantization and multi-GPU support.
AI Snips
Chapters
Books
Transcript
Episode notes
Elixir Helps ML Infrastructure Not Raw Compute
- Elixir's concurrency and BEAM strengths help deployment and app-level workflows, but heavy ML compute speedups come from native libraries and compilers (C/C++/XLA).
- So Elixir shines in integrating ML into apps and building robust data pipelines rather than raw tensor speed.
Explore Data First With Explorer
- Use Explorer for data exploration and visualization early to understand distributions, outliers, and whether ML is needed.
- Compare simple models (linear, tree) and neural nets, evaluate accuracy and deployment cost before selecting a model.
Dataset Bias Creates Hidden Failure Modes
- Models have failure modes tied to dataset bias; inspecting examples reveals spurious correlations (e.g., dumbbells always shown with hands).
- Continuous monitoring and representative data collection prevent these blind spots after deployment.
