Towards Typesafe Deep Learning in Scala

on YouTube ↗35m

" summary

Nexus is a prototype Scala deep‑learning engine that encodes tensor shapes and axis semantics into types so axes of tensors are typed statically and operations either compile or fail. It uses Shapeless HLists, Scala implicits as compiling-as-proofs, and GADTs to enforce rules like matrix multiplication only when inner axes match and sum‑over‑axis producing a removed axis at compile time. Backend bridges target native C++/CUDA (PyTorch/MXNet/TensorFlow), GPU support is preliminary, automatic batching and GC for device memory are planned, and models can shrink from ~1000 lines to under 20 in examples.

AI-written summary · about

" chapters (8)

1. 0:00Introduction and motivation
2. 5:25Problems with Python deep learning
3. 10:20Type-safe tensors in Scala
4. 19:37Polymorphic symbolic functions
5. 26:52Native CUDA integration plans
6. 28:43Dynamic neural architectures
7. 31:15Lazy and eager evaluation
8. 33:40Q&A

" related talks

• 54mWhy Netflix ❤'s Scala for Machine LearningJeremy Smith, Aish · Scala Days Lausanne 2019
• 34mDifferentiable Functional ProgrammingNoel Welsh · Scala Days Berlin 2018
• 32mTowards new deep learning abstractions on top of exist frameworksTristan Webb · Strange Loop 2017
• 23mDeep dive in Nx BackendsPaulo Valente · Code BEAM America 2022
• 40mScala and Spark in the New Deep Learning EraEmiliano Martinez Sanchez · Scala Days 2023 Madrid
• 37mMonadic Deep LearningYang Bo, Xiaolei Wang · Strange Loop 2017
• 31mScoring ONNX ML Models with ScalaOscar Boykin · Scale By The Bay 2019
• 24mKeynote: Next-generation frameworks for Large-scale AIAnima Anandkumar · Scale By the Bay 2020