1

Integration of Deeplearning4j withApache Apex

Ambarish PandePriyanka Gugale Devraj Baheti

Priya HedaPriyanka Kekane

2

• What is Deep Learning?

• Why Deeplearning4j?

• Using Deep Learning with Apex

• Architecture of the integration

• Challenges faced during the integration

• Demo

Agenda

3

• Deep learning is a branch of machine learning based on a set of algorithms that attempt to model high level abstractions in data.

• Deep learning eliminates the need for feature engineering.• Effectively works on unsupervised data.

Deep Learning

4

• Deep learning uses deep neural networks to model the high level abstractions in data.

• Deep neural networks are neural networks with more than one hidden layer.

Deep Learning

5

● Colorization of black and white images• Object Classification in Photographs• Automatic Game playing• Image Caption Generation• Handwriting Recognition• Automatic Machine Translation• Adding sound to silent movies

Applications of Deep Learning

6

● An Open Source Deep Learning library( released under Apache 2.0 license)● DL4J is Distributed● Written for Java and Scala● Integrated with Hadoop● Skymind is its commercial support arm● The Neural Net platform Dl4j provides various neural networks like Long

Short-Term Memory units, Convolutional Neural Networks for image processing, Deep AutoEncoder, Restricted Boltzmann Machine, Recurrent Nets, Denoising Autoencoders etc.

Deeplearning4j

7

Deeplearning4j

8

● Training Deep Learning models on single processor is extremely slow.

● Dl4j works with multi CPU and multi GPU systems.

● This integration will enhance the implementation of deep learning models in

distributed and stream processing environments.

Using Deep Learning with Apex

9

• We achieve distributed training of neural networks using Data Parallelism.• In data parallelism, different machines have a complete copy of the model,

each machine simply gets a different portion of data.

Architecture of the integration

10

● We use a method called Parameter Averaging to combine and synchronize models trained on different machines.

Architecture of the integration

11

DAG for the integration

12

● The Iris flower data set or Fisher's Iris data set is a multivariate data set.● The data set consists of 150 samples from each of three species of Iris (Iris

setosa, Iris virginica and Iris versicolor). Four features were measured from each sample: the length and the width of the sepals and petals, in centimetres.

Dataset used

Iris versicolor Iris setosa Iris virginica

13

DEMO

14

15

• Had to change default packaging of Apex Application

• We used Maven Shade plugin for packaging the app.

• Certain components of Nd4j are incompatible with KryoSerializer.

• We are using Java Serializer for those components.

Challenges faced during Integration

16

Thank You!