Convolutional Neural Networks at scale in
summary
Implemented convolutional neural nets inside Spark MLlib that run convolutions, max‑pooling and dense layers with a Keras-style Scala API, using distributed backprop via SGD with adaptive gradients (momentum, RMSProp, Adam). Cluster training averages worker gradients synchronously or asynchronously, giving the iterative workloads 10–100× speedups versus disk‑based MapReduce. Supports transfer learning from ImageNet features, shows faster convergence and better stability, and roadmaps include GPU acceleration, Python bindings and Keras integration.
AI-written summary · about
chapters (13)
- 0:00Start
- 0:28Jeremy Nixon
- 3:18Deep Learning in MLlib
- 4:29Structure
- 6:10Framing Convolutional Neural Networks
- 6:21Structural Assumptions
- 12:51Applications
- 15:47Automated Feature Engineering
- 17:00Learning Representations
- 17:17MLlib Flexible Deep Learning API
- 20:00Distributed Optimization
- 20:13Performance
- 20:29Future Work
related talks
22mBringing Deep Learning into the Big Data Analytics
39mMachine Learning with Scala on Spark
17mMachine Learning at the Speed of Light with Spark
40mScala and Spark in the New Deep Learning Era
23mRethinking scalable machine learning with Spark ecosystem
34mDeep Learning at Scale
31mFrom Pandas to Spark
39mA full Machine learning pipeline in Scikit-learn vs in scala-Spark: pros and cons