Numenta Logo

A Mathematical Formalization of HTM’s Spatial Pooler

James MnatzaganianJames MnatzaganianCommunity Contributor
A Mathematical Formalization of HTM’s Spatial Pooler

Those of you subscribing to the nupic-theory mailing list are aware that a new research paper describing a mathematical model for the spatial pooler (SP) has emerged. Many of us have asked “What is the math behind the SP?” or “How can I use the SP for machine learning”. The goal of this paper is to address those very questions, bridging the gap between HTM and the machine learning community.

This work is part of a much larger body of work being conducted by the Rochester Institute of Technology’s (RIT’s) NanoComputing Research Lab. Our lab is specifically focused on designing energy efficient hardware circuits and architectures that are biologically inspired. One of our key strengths is the utilization of emerging technologies such as memristors to design energy efficient neuromorphic systems.

So why is a hardware group interested in mathematical models?! To properly design hardware, it is imperative to build on a key mathematical model of the overarching system. If we are able to do so, we could potentially produce a hardware design of HTM that not only has higher performance than a software-based solution, but also consumes much less power. We are currently working on such a system, so please follow our work. ☺

It is widely known that the SP is similar to a self-organizing map (SOM). Our paper takes that idea farther, by showing that the primary learning mechanism consists of a component that is very similar to competitive learning. In fact, we posit that the SP is not “a” competitive learning network, but rather many competitive learning networks. In this construct, each column acts as a competitive learning network, with the columns’ set of inputs determined in a manner similar to attribute bagging. The final set of active columns is then determined by a voting scheme. This process can be used to explain the reasoning behind the permanence selection, but it still leaves the reasoning behind the permanence update method.

Many people have posited that the permanence update rule is similar to Hebbian learning, since winning columns have their active synapses positively reinforced. While this is a nice connection, it is not sufficient for a full mathematical model. We demonstrate how the permanence update may be modeled as an optimization problem, through the use of a maximum-likelihood estimator (MLE). Using this approach, it is possible to choose permanence increment and decrement amounts best suited for the application.

In addition to the discussion on the primary learning mechanism, we show how the SP can be used to perform feature mapping and dimensionality reduction. A method is also provided for reconstructing the SP’s representation back in the input space. We additionally provide full equations for every aspect of the algorithm (including the boosting operations), optimized for matrix computations. A discussion concerning the initialization of the network is also included.

For those of you more interested in the overall system design, we have open-sourced our code (MIT License). You may download the full source from my GitHub and you may browse the fully generated API on my personal website. If you are familiar with machine learning in Python you will be happy to know that the implementation uses the scikit-learn interface. Our implementation includes support for using scikit-learn’s build in cross-validation (CV) suite. We also have sample code on how to use this as a single process or as many processes locally or as a plethora of process on a cluster.

It is our wish that this work helps the community to further advance the understanding and application of HTM. We are continuing this work by studying the temporal memory portion of HTM.

As with all work, it is important to note any contributions. K. Gomez of Seagate Technology [10] and J. Hawkins, S. Ahmad, and Y. Cui of Numenta provided feedback on this work. Many members of the NanoComputing Research Lab (namely A. Hartung and C. Merkel) also provided critical feedback. Lastly, RIT’s research computing provided the infrastructure to allow us to perform our simulations.

James MnatzaganianJames MnatzaganianCommunity Contributor