Neural network implementations in Java

If by "scalable" you mean speed: according to netbeans zone:

Two major open source neural network projects for the Java platform, Encog and Neuroph, have announced collaboration [...]. The basic idea for collaboration is to provide the best of both worlds: Encog’s speed [...] and Neuroph’s [...] easy to use neuron-based interface on top for advanced neural network research. This will be accomplished by creating a high performance Encog kernel which will also be used by Neuroph.

Alternatively: have a look at Weka, which is extensible and already contains a lot of ML implementations.

Edit ("by scalable I mean ability to work with huge datasets"): for large datasets, maybe a framework like Apache Mahout is suitable for you: Apache Mahout's goal is to build scalable machine learning libraries. With scalable we mean: Scalable to reasonably large data sets...Scalable to support your business case etc

Also: SQL Server 2005+ supports DM and allows for further ML algorithms to be deployed inside it. However, is is a .NET-centric world. Theoretically, by using IKVM you could develop Java packages and deploy them as .NET assemblies.