dl.md

#Deep Learning Tutorial

The purpose of this tutorial is to walk new users through Deep Learning using H2O Flow.

Those who have never used H2O before should refer to Getting Started for additional instructions on how to run H2O Flow.

For more details on the math behind H2O's implementation of Deep Learning, refer to Deep Learning Data Science.

For tips on improving the performance and results of your Deep Learning model, refer to our Definintive Performance Tuning Guide for Deep Learning.

###Using Deep Learning

H2O’s Deep Learning functionalities include:

• purely supervised training protocol for regression and classification tasks
• fast and memory-efficient Java implementations based on columnar compression and fine- grain Map/Reduce
• multi-threaded and distributed parallel computation to be run on either a single node or a multi-node cluster
• fully automatic per-neuron adaptive learning rate for fast convergence
• optional specification of learning rate, annealing and momentum options
• regularization options include L1, L2, dropout, Hogwild! and model averaging to prevent model overfitting
• elegant web interface or fully scriptable R API from H2O CRAN package
• grid search for hyperparameter optimization and model selection
• model checkpointing for reduced run times and model tuning
• automatic data pre and post-processing for categorical and numerical data
• automatic imputation of missing values
• automatic tuning of communication vs computation for best performance
• model export in plain java code for deployment in production environments
• additional expert parameters for model tuning
• deep autoencoders for unsupervised feature learning and anomaly detection capabilities

Getting Started

This tutorial uses the publicly available MNIST data set of hand-written digits, where each row contains the 28^(2)=784 raw gray-scale pixel values from 0 to 255 of the digitized digits (0 to 9).

To further explore H2O's capabilities, some publicly available data sets can be found on our website.

####Importing Data Before creating a model, import the data into H2O:

0. Click the Assist Me! button in the Help tab in the sidebar on the right side of the page.

0. Click the importFiles link and enter the file path to the training dataset in the Search entry field, or drag and drop the file onto the Search entry field.

NOTE: Deep Learning requires a training dataset and a validation (or testing) dataset. Make sure to upload and parse both datasets before creating a model.

0. Click the Add all link to add the file to the import queue, then click the Import button.

####Parsing Data

Now, parse the imported data:

0. Click the Parse these files... button. Note: The default options typically do not need to be changed unless the data does not parse correctly.
1. From the drop-down Parser list, select the file type of the data set (Auto, XLS, CSV, or SVMLight).
2. If the data uses a separator, select it from the drop-down Separator list.
3. If the data uses a column header as the first row, select the First row contains column names radio button. If the first row contains data, select the First row contains data radio button. You can also select the Auto radio button to have H2O automatically determine if the first row of the dataset contains the column names or data.
4. If the data uses apostrophes ( ' - also known as single quotes), check the Enable single quotes as a field quotation character checkbox.
5. To delete the imported dataset after the parse is complete, check the Delete on done checkbox.

NOTE: In general, we recommend enabling this option. Retaining data requires memory resources, but does not aid in modeling because unparsed data cannot be used by H2O.

0. Review the data in the Data Preview section, then click the Parse button.

NOTE: Make sure the parse is complete by clicking the View Job button and confirming progress is 100% before continuing to the next step, model building. For small datasets, this should only take a few seconds, but larger datasets take longer to parse.

##Building a Model

0. Once data are parsed, click the View button, then click the Build Model button.
1. Select Deep Learning from the drop-down Select an algorithm menu, then click the Build model button.
2. If the parsed training data is not already listed in the Training_frame drop-down list, select it.

Note: If the Ignore_const_col checkbox is checked, a list of the excluded columns displays below the Training_frame drop-down list.

0. From the drop-down Validation_frame list, select the parsed testing (validation) data.
1. From the Ignored_columns section, select the columns to ignore in the Available area to move them to the Selected area. For this example, do not select any columns.
2. From the drop-down Response list, select the last column (C785).
3. From the drop-down Activation list, select the activation function (for this example, select Tanh).
4. In the Hidden field, specify the hidden layer sizes (for this example, enter 50,50).
5. In the Epochs field, enter the number of times to iterate the dataset (for this example, enter 0.1).
6. Click the Build Model button.

##Results

To view the results, click the View button. The output for the Deep Learning model displays the scoring history, the training metrics, and the status of the neuron layers.

For more details, click the Inspect button.

Select the appropriate link to view details for:

• Neuron layer status
• Scoring history
• Training metrics
• Training metrics confusion matrix
• Validation metrics
• Validation metrics confusion matrix

The training metrics confusion matrix is shown below.