1. Leave out all the noise features before training the neural network. Remove noise features at the schema transformation stage:

TransformProcess transformProcess = new TransformProcess.Builder(schema)
 .removeColumns("RowNumber","CustomerId","Surname")
 .build();    

2. Identify the missing values using the DataVec analysis API:

DataQualityAnalysis analysis = AnalyzeLocal.analyzeQuality(schema,recordReader);
 System.out.println(analysis); 

3. Remove null values using a schema transformation:

Condition condition = new NullWritableColumnCondition("columnName");
 TransformProcess transformProcess = new TransformProcess.Builder(schema)
   .conditionalReplaceValueTransform("columnName",new IntWritable(0),condition)
 .build();

4. Remove NaN values using a schema transformation:

Condition condition = new NaNColumnCondition("columnName");
 TransformProcess transformProcess = new TransformProcess.Builder(schema)
   .conditionalReplaceValueTransform("columnName",new IntWritable(0),condition)
 .build();