Numenta Logo
Blog Receives New Network API

David RayDavid RayCommunity Contributor Receives New Network API

Greetings Earthlings! (…and otherwise affiliated) :-) We’re [ insert understatement here ]excited?” to announce the completion of the Network API (NAPI) for the Java™ port of NuPIC. This addition to will usher in a whole new realm of possibilities for application integration and for streaming data applications.

Stream of consciousness you say? Maybe someday… :-) For now let us be content with easy integration of Numenta’s learning, prediction, and anomaly detection algorithms into new forward-thinking Java applications; distributed data applications; and enterprise ecosystems.

Until now, Java users of NuPIC had to be content with piecing together individual algorithmic components by supplying their own duct-tape, stitching and glue. This is no longer necessary, as the NAPI is very robust; has lots of features, and most importantly - is hypnotizing-ly easy to work with!

Here’s an example of the code it takes to get a full featured network up and running:

Parameters p = NetworkDemoHarness.getParameters();
p = p.union(NetworkDemoHarness.getNetworkDemoTestEncoderParams());

Network network = Network.create("Network API Demo", p)
  .add(Network.createRegion("Region 1")
  .add(Network.createLayer("Layer 2/3", p)
  .alterParameter(KEY.AUTO_CLASSIFY, Boolean.TRUE)
  .add(new TemporalMemory())
  .add(new SpatialPooler())
  .add(Sensor.create(FileSensor::create, SensorParams.create(
    Keys::path, "", ResourceLocator.path("rec-center-hourly.csv"))))));


…and that’s it!

For less “ear-chattering” and more “quick-starting”, see the Quick Start Guide & Docs.

Decided to stick around eh?

Well, since you’re here, let’s talk more about how the NAPI is designed and a bit more about its features.

In, the hierarchical entities seek to provide corollaries to actual biological structures. A Network can be thought of like a single neocortex, while a Region can be thought of as a location in the neocortex that is sensitized to a particular function, behavior or memory. Lastly, Layers are metaphors for actual layers in the neocortex such as the 6 layers identified by neurological researchers. Thus, in, the Layer is the container for the Connections object and eventually will hold the column/cell matrix that the Connections object now holds. (Hint!: Work to be done here). The SpatialPooler and TemporalMemory describe how connections are formed and their associations, and algorithmic data flow - but are not physical members of a hierarchy - they describe how data flows.

At the top level there is the Network. A Network is a container of Regions, which in turn contain Layers, which in turn contain Columns, Cells, Dendrites and Synapses - and are acted upon by the different algorithms (i.e. SP, TM etc.). A Network can be both listened to and it can have data submitted to it. In fact, this is its normal usage pattern.

Regions are collections of Layers. While they are not an optional construct within a Network, they also don’t do all that much. Regions are best thought of as “groups” of algorithmic hierarchies that may in the future, specialize on types of knowledge or types of behavior or other functionality. They are here mostly to provide a corollary to Regions in the neocortex - but mostly they are meant to identify and distinguish layers of functionality specialized for a particular purpose.

The last level in the hierarchy is the Layer. Layers contain the column/cell matrix (Connections) and house all the algorithmic and functional components. This is where the heart of the NAPI beats. Layers house most of the functionality - automatically forming computational units implemented via RxJava Observables which know how to connect one algorithm to another, transforming its data in the process in order to input data to the next succeeding algorithm (whatever it may be) in the format that that algorithm expects. Layers emit Inference objects when subscribed to (either via their Layer.subscribe() method or by retrieving their Observable<Inference> object, and calling Observable.subscribe() on it); and may have data directly submitted to them via the onNext() method of their Observable. To retrieve their Observable, simply call Layer.observe() which returns an Observable<Inference> object. While the Layer object is flexible enough to function this way, the typical usage pattern will be to create a Network object (with a Sensor connected at the bottom), start it; then subscribe to it via its Observable (Network.observe()), in order to receive its emitted Inference objects. All hierarchy objects may be “listened” to by calling (Network.observe(), Region.observe(), or Layer.observe() / Layer.subscribe()).

The Inference object emitted contains a vast amount of information such as the predicted columns, the previously predicted columns, the activated columns etc. In addition data may be retrieved from a given Layer following each cycle also.

Usage (A bit of FRP)

First, as you can see above, entire networks can be constructed using the "fluent" style of declaration (all one method call chain). This reduces boilerplate typing and clarifies the hierarchical structure of the api’s constructs.

Secondly, the whole framework is designed with Streaming data and Asynchronous usage in mind. The network’s Sensors were implemented using the Java 8 Streams API and will be totally asynchronous ready when work on the various Encoders is completed to make them thread safe.’s Stream usage has been customized to include “batching” of input - speeding up parallel processing by up to 30 or 40%. The BatchedCsvStreamBenchmark test class can be run as a demonstration. (Your mileage may vary.)

Finally, the RxJava Reactive Extensions library was used throughout to simplify the connection of algorithms within a Layer.

Think about it…

Each Layer can contain one or more components, and it is unknown which ones will exist in any given Layer. Next, these components all output different data and take in different data in different combinations. So a framework that allows for all of this must be adaptable and able to transform data based on the dynamic existence or non-existence of any particular algorithm or component!

Additionally, because at all levels (except the bottom input level), Inference objects are passed between components, layers and regions; RxJava Observables can be obtained at any point in the hierarchy, and subscribed to to receive output data --or-- they can be combined, operated upon and manipulated in Map-Reduce form! How exciting is that?

This means that you can take output from any point and operate on it (combine, zip, merge, delay etc.) in your own applications using RxJava idioms!

A bit about input / output (FEED ME!)

This is discussed in more detail in the wiki linked below, but we just wanted to touch on this briefly.

Networks can be fed either by connecting them to receive streaming data "automatically" by declaring a FileSensor or URLSensor. Additionally you can call Layer.compute() directly with either an int[], String[] (not fully implemented), Map (like an encoder mapping), or ManualInput object. Lastly, you can use an ObservableSensor to either feed in data manually/programmatically –or– connect to other Observable emitters within your applications!

ManualInput objects inherit from Inference (the interface received on output and passed between layers, regions, and other components)

At the other end, Networks can be subscribed to in order to receive data "out of the top" (And so can Regions, and Layers - see Region.observe() and Layer.observe() ).

Here’s a brief look at the Observable interface’s simplicity:

Observable<Inference> o = network.observe();
// This Observable can be used a million different ways!
o.subscribe(new Subscriber<Inference>() {
  @Override public void onCompleted() {
    // Do finalization work here if desired

  @Override public void onError(Throwable e) {
    // Error handling here

  @Override public void onNext(Inference i) {
    // Called on every cycle of the Network - work is done here

Where do I fit in?

Now that the ground work has been laid, there is a TON of things to do! If you would like to get your name in bright lights, or simply help out with drop a line to the hackers mailing list and offer to help - all are welcome (at any level!).

For more in depth information (Work in progress), start here with the NAPI - In-Depth Component Overview; which leads into a tutorial series of wiki pages and docs.

To dig right in, look at the Quick Start Guide


I’d like to mention that I’ve had a lot of help from my team members and the leadership at with organizing my work, and also being generously granted the time to work on this project! (I have such a great job!) - Thank you!

Happy Hacking!


David RayDavid RayCommunity Contributor