In this video:
- Romeo Kienzler, Chief Data Scientist, IBM
Romeo presents at GPU Technology Conference, and talks about deep learning and applying it to IoT sensor data, which is structured data and time-series data, and using any of the open source deep learning frameworks for analyzing the IoT data.
After giving a brief introduction to deep learning and LSTM networks (read his tutorial for more details), Romeo discusses 4 ways of doing parallelization and introduces the deep learning frameworks for Apache Spark and how to turn on GPU and achieve parallelization:
- Inter-model parallelism (TensorFrames)
- Data parallelism (DeepLearning4J, Apache SystemML, TensorSpark, CaffeOnSpark)
- Intra-model parallelism (ND4J, Apache SystemML)
- Pipelined parallelism
He finally discusses different messaging protocols, including broadcasts, GRPC, and Open MPI, and finally IBM Spectrum MPI, giving IBM the fastest AI system because of connection between CPU and GPU.
Resources for you
- Try Romeo’s tutorial on building an anomaly detector using DeepLearning4J
- Try Romeo’s tutorial on building an anomaly detector using Apache SystemML
- Watch a replay of this live coding event where Romeo creates a deep learning anomaly detector using Apache SystemML
- Learn more about the different models for machine learning in this developerWorks article.