This document discusses DeepLeanring4j, a framework for data parallel deep learning on Spark. It provides an overview of the current landscape of deep learning tools, and proposes a solution using Javacpp, libnd4j, and SKIL (Skymind Intelligence Layer) to leverage Spark and the JVM ecosystem while allowing for heterogeneous compute. Key aspects include efficient access to image and audio data, an ND4j library for numerical compute, deployment via Juju, and an ETL pipeline interface in Canova. The goal is to build on the JVM's strengths while addressing its limitations for numerical compute through hardware acceleration.

1. DeepLeanring4j
Data Parallel deep learning on spark

2. The JVM is too slow for numerical compute
Great at network I/O and data access
Great streaming infrastructure
Hardware accel required
Spark - Data Access Layer.
Cuda - Compute layer

3. Current Landscape
Spark assumes columnar data
Binary (audio/images) is becoming more important
HDFS is great for storing blobs
SQL doesn’t work for pixels and audio frames
The ingredients are here for something great

4. The solution
Javacpp (cython for java)
64 bit pointers for efficient contiguous access of image and audio data
Leverage java’s distributed systems ecosystem
Add new numerical compute layer (libnd4j)
Allow for heterogeneous compute
Off heap memory
Easy deployment
Data pipelines as a first concern

5. SKIL (Skymind Intelligence Layer)

6. JavaCpp
Auto generate JNI bindings for C++ by parsing classes
Allows for easy maintenance and deployment of c++ binaries in java
Write efficient ETL pipelines for images via opencv (javacv)
Integrate other c++ deep learning frameworks (tensorflow,caffe,..)
Allows for productionization of fast (but academic) C++ code using java
(kafka,spark) for ETL
64 bit pointers (wasn’t possible before)

7. “Actual” Streaming frameworks
Kafka
Flink
Spark Streaming
Apex

9. Nd4j
Heterogenous codebase
Supports cuda, x86 and soon (power)
Shared indexing logic for writing ndarray routines
Memory management in java (even cuda memory!)
Openmp on cpu + routines for common things such as reduce
Pinned memory and async operations
JIT allocation
Spark friendly (runs on multiple threads and devices)

10. Deployment
Juju
Runs as spark job
Easy to embed in production

11. Canova
One interface for ETL
Integrates with spark
Easy to extend to write your own custom data pipelines
One interface for generating NDArrays

12. Conclusion
Built to be friendly to the JVM ecosystem
Allows java to do what its good at
Numpy in java means easy to port things like scikit learn
Data Parallel means commodity hardware JVM assumes works

13. Future
Model Parallelism
Opencl
Sparse support
Reinforcement learning

14. Questions?
help@skymind.io