Nupic Node

This sample project is distributed together with the NupicNode through a separate NupicModule repository, because of different licensing (GPL v3). The sample project file is located in the BrainSimulatorExample folder.

The NupicModule contains the Numenta's nupic.core C++ codes, needed external libraries and a C# wrapper, that allows to run the nupic.core functions from inside of the Brain Simulator node (MyNupicNode.cs, MyNupicTasks.cs). We made use of the Numenta's open source project in order to have the official Cortical Learning Algorithms (CLA) implementation, so we can have a direct comparison and we can easily utilize the future changes from the Nupic community.

Installation of the Nupic Node

To start to use the NupicModule you have to:

CLA MNIST Prediction Example

Brain: NupicModule\BrainSimulatorExample\CLA_mnist_prediction.brain

If the NupicModule is set up correctly, you can now load the NupicModule\BrainSimulatorExample\CLA_mnist_prediction.brain project, run the simulation and you shall see the following workspace: The project is set up for the MNIST hand-written numbers recognition and series prediction using the CLA algorithm implemented in the NupicNode. The available inputs are the standard 28 * 28 MNIST bitmap coded with float numbers (0 to 1) and the Label, which is the integer number that is represented on the bitmap. First, we do some pre-processing with the nodes Prescaling and ProcessedMnistInput in order to have the bitmap in a correct format for the NupicNode input - it expects a binary (0,1) vector. We connect the processed MNIST Bitmap and the input Label to the NupicNode's inputs Input and Label. There are several outputs from the NupicNode:

In the example brain file, we are predicting only a single time step that is 1 step in the future (this is set up in the NupicNode's property PredictedStepsList - it can contain a list of more comma separated time step values and predict several of them simultaneously). This value - the NupicNode's prediction is visualised in the top left observer. Below it is shown the actual input into the NupicNode. The input is being sent in a sequential way, every digit from 0 to 9 is sent one after another (after 9 it wraps up to 0 again), so the correct prediction should be exactly by 1 higher than the one in the current input.

NupicNode Parameters

The NupicNode has plenty of parameters to fine tune, which is very impractical to do manually. The actual parameters used in this example were obtained using the swarming optimization algorithm, present in the Nupic framework (currently working only in Linux/MacOS). The core CLA parameters were exported into the nupic_params_from_swarm.txt file and are automatically loaded and set to the NupicNode through its property PropertiesFileName. The description of the swarming process is beyond the scope of this document, but you can find it on Nupic's Github pages.

Results

The bottom right observer CLA_1_StepPrediction - ClassifierOutput depicts the probability of different buckets. We can see, that sometimes the classifier assigns high probabilities to more buckets, i.e. when the current input is 9, it can predict both, 0 and 8 with a high probability, because they are similar from the NupicNode's point of view. The rightmost node is a NodeGroup that contains just several nodes for the calculation of error. The sliding average of the error over the last 100 time steps is plotted in the largest TimePlotObserver: AverageErrorOverLast100Steps. After starting the simulation, we can see that the error peaks somewhere around 100 steps, that means that the NupicNode needs to see about 100 input values (each digit 10 times) until it learns to recognize and predict them. The top right observer prints the current average error value and the one below it prints the total number of misses since the start of the simulation. The average error rate of prediction is around 7% in this experiment.

Discussion, Input Encoding

When we look on the setting of the MNISTWorld, into the parameter ImagesCnt of the task Init MNIST World, we can see the total number of input images is 100 (10 for each digit). The algorithm doesn't work very well with larger number. This shows the largest disadvantage of this setting of the experiment. The NupicNode with these parameters (mainly limited by the number of columns) and this format of input is not able to learn and more importantly generalize larger number of different shapes of digits. The main contributor to this is probably the encoding of the input MNIST bitmaps, which is not very appropriate for the CLA algorithm. The CLA requires binary input vectors with Sparse Distributed Representation (SDR) properties, e.g. containing much lower number of 1s than 0s (around 2%) and the meaning highly correlated with the position of 1s in the vector. The testing and creation of better encodings will be the next step in our research. Currently there is only one simple encoder present in the Brain Simulator BasicNodes - the ScalarToSDRNode which encodes scalar numbers into SDRs, with which you can experiment. Nevertheless, the example shows that the CLA can very quickly learn to accurately recognize and predict the patterns in a time series.