Monadic Deep Learning

on YouTube ↗37m

" summary

Claim: DeepLearning.scala 2.0 implements dynamic neural networks by modelling layers as composable monadic computations and exposing an monad-based DSL plus an Applicative type class so branches can run in parallel. Implementation specifics: convert layers to differentiable computations (flatMap, pair combinator), mark tasks for parallel GPU/CPU execution, plugin-style hyperparameter and layer injection, and a planned element-wise higher‑order operations path to solve ND4J’s current GPU limitation. Performance tradeoffs and examples of conditional nets and parallel scheduling are shown.

AI-written summary · about

" chapters (12)

1. 0:00Introduction to the topic
2. 1:15Challenges in deep learning
3. 3:45Scala and deep learning
4. 6:03Programming environment
5. 8:23Neural network implementation
6. 12:04Higher-order functions
7. 14:05Functional programming patterns
8. 17:51Conditional neural networks
9. 21:03Monads and performance
10. 24:04Parallel execution on GPU
11. 27:53Managing hyperparameters
12. 33:14Future work and conclusion

" related talks

• 34mDifferentiable Functional ProgrammingNoel Welsh · Scala Days Berlin 2018
• 39mDifferentiable Functional ProgrammingNoel Welsh · Scala Days New York 2018
• 32mStructured Deep Learning with Probabilistic Neural ProgJayant Krishnamurthy · Scale By The Bay 2018
• 37mNeural Networks are Programs, TooMarisa Kirisame · LambdaConf 2017
• 29mDeep Learning Needs ClojureCarin Meier · Clojure/conj 2017
• 48mDynamic Neural Networks with DyNetYoav Goldberg · PyCon Israel 2017
• 51mDeep Learning - What is it and What It Can Do For YouDiogo Moitinho de Almeida · GOTO Copenhagen 2016
• 30mConstrained Monotonic Neural NetworksDavor Runje · DSC Croatia 2022