Frequently Asked Questions


Q:

What should I read to get prepared for working with Numenta technology?

A:

Hierarchical Temporal Memory is new, so there's not a lot of HTM-related information available. For starters, we recommend that you read On Intelligence, HTM Concepts whitepaper, and our getting started guide. Additionally, there are a lot of interesting technologies that Numenta uses to implement HTMs. We've created an education page, Other Reference Materials, to help you come up to speed quickly. Additionally, we've created a Wiki page for the Numenta Community to pitch in and help suggest other materials.



We have a Getting Started topic in the Education forum to help you learn about HTM.


Q:

Where can I download your product?

A:

Downloading the Numenta Platform for Intelligent Computing (NuPIC) is quite simple. Use one of the following links to download the software:


Q:

Why is this a Research Release?

A:

The Research Release is a no-charge version of the Numenta Platform for experimentation purposes. Our goal is to help kick-start an active development community around HTMs. We welcome your participation in this community, but we must warn you up front: it is a Research Release because the platform tools are new, the platform is young, and the concepts are sometimes difficult. To make real progress, you can expect to dedicate significant time to learn the material, and then to applying it to your problem domain. We hope that you will be an active participant, sharing your work, reporting bugs, and contributing to our forums.

Please see CEO Donna Dubinsky's note announcing the Research Release for more information.


Q:

Can you show that HTM Networks are better than other technologies?

A:

We don't have any benchmarks of HTMs at this point. However we have some documentation that compares our technology with competing solutions. Read more about how we compare in our HTM Comparison document.

Q:

Where can I find the mathematical formulas for your HTM algorithms?

A:

For our first version of our algorithms, Zeta1, we have a document that completely describes our algorithm formulas. These can be found in our software release in the /share/doc directory, under the name 'Zeta1AlgorithmsReference.pdf'.

Q:

When is your commercial release? What will be your pricing strategy?

A:

We expect to create a commercial release of our platform some time in the future, although no dates have been planned yet. The exact pricing structure has not been determined, but please see the Welcome Blog for a discussion on pricing philosophy and for a high level description of each license as it relates to commercial deployment

Q:

What are your investment opportunities?

A:

Numenta is a private company in the development phase. It is currently funded by its founders and board members. The company is not looking for other investors at this time. If you want to be updated on Numenta's progress and status, we recommend that you sign up for our newsletter.

Q:

Your licenses have recently changed. Can you describe all of the changes?

A:

We have written an entire page devoted to the changes, and intentions, of our new licenses. Please visit our Licensing Page. However, don't rely on this blog for precise information. You still have to read the licenses in their entirety.

Q:

I noticed that you have two downloads: General Research and Academic Research. What are the differences?

A:

We have two different licenses divided by audience. One license, the academic license, is for students, faculty and research staff at accredited educational institutions. The other license, general research license, is for everybody else. For more information, please visit our Licensing Page.

Q:

I downloaded your platform, and it said a license key was sent to me, but one isn't in my mailbox. What do I do?

A:

Our system tries to send out the license key within five minutes of your downloading of the platform. If you don't get your license key, try the following:

  • 1. Check your junkmail folder. Often, over-zealous junk mail programs may put our emails into a junk folder. Check the junk folder before moving forward.

  • 2. Check your profile. Go to your Numenta Profile, and check your email settings. Perhaps you've entered it incorrectly. If so, specify the correct email address, and re-download the software.

  • 3. Re-download the software. No system is completely fault tolerant. Perhaps the system didn't send out the key as requested. Re-download the software and the key should be re-mailed to you.

  • 4. Email support. Still no luck? Send an email to key@numenta.com and request a new key. Be sure to include your username, name, and email address.


    • Q:

    I received a license key, but what do I do with it?

    A:

    The Numenta Platform requires the presence of a license file in order to execute any functionality. After you've registered on the Numenta website, and downloaded the platform, you will receive an email with your license configuration file, prefaced with your username. For example, if your username is NumentaPhil, the license file you receive will be NumentaPhil-license.cfg.

    For Macintosh or Linux:

    • Copy the license file into $HOME/.nta, as in:
    •   cp Numenta-Phil.license.cfg ~/.nta/

    Note: The name of the directory is .nta and not nta.

    For Windows:

    • Copy the license file into your documents folder. On Windows XP this is the My Documents folder, and the Documents folder on Windows Vista.

    Q:

    When I login the website returns to the main page and it says I'm not logged in. What happened?

    A:

    Our website relies on cookies in order to keep you logged in. If you don't have cookies enabled in your browser, then you will not be able to login to the system. Edit your browser preferences to enable cookies at the numenta.com website.

    Q:

    Sometimes your website says I'm logged in but then requires me to login again. Why is that?

    A:

    If you are logged in to the website and inactive for over 24 minutes, you will be logged off. At this time, however, our website does not refresh your login status, but this will be remedied shortly.

    Q:

    I am connected to the Internet but the Pictures Demonstration Application is saying it can't connect. What can I do?

    A:

    It is possible that anti-virus and/or firewall software installed on your Windows® machine is blocking the attempted connections from Pictures Demo to the Numenta HTM cluster. Some security software forbids programs other than browsers (such as Internet Explorer) from initiating outbound web connections. You may be able to resolve the issue by modifying your security settings to allow the program 'RunPictures.exe' to make outbound HTTP requests.

    Q:

    The Pictures Demonstration Application is not responding when I click 'Recognize Picture'. What can I do?

    A:

    The Windows® version of the Pictures Demo uses a cluster of HTM servers maintained by Numenta in California. This cluster can support only a finite number of simultaneous picture recognition requests. If your Pictures Demo program is timing out, not responding, or failing to connect to the servers, then it is likely that the Numenta HTM servers are too busy responding to other users. You might consider running the Pictures Demo again at a later time.

    Q:

    I am not able to find the software prerequisites you recommend. What can I do?

    A:

    To help simplify the installation process, Numenta has created a Software Prerequisites page for your use.

    Q:

    Does Numenta make the open source tools they use available to the general public?

    A:

    Numenta has put together an external source release for your use. Please visit the External Software Source page for more information.

    Q:

    How do I determine the correct network topology?

    A:

    Your network topology must be appropriate for the type of data you are feeding in and the "world" which you are trying to model. This is an active area of research for HTMs. You can use the existing examples as guidelines, though you likely will need to experiment with a few topologies. The Pictures example contains a self-contained NetExplorer experiment for trying out a few different topologies. Ultimately the best topology depends on the structure of your data and the underlying causes. For an ongoing conversation about this, please see the Network Topology Wiki page.

    Q:

    How much data do I need?

    A:

    You need sufficient data for each level of the hierarchy to gather spatial and first order temporal statistics (correlations between successive inputs). In general the more data you have the better the statistical model inside your HTM. The HTM still will work with very little data but overall accuracy may be low. The Pictures example contains a self-contained NetExplorer experiment that plots the effect of the number of training examples on the end accuracy.

    Q:

    What is the difference between supervised and unsupervised learning?

    A:

    If you train an HTM network using supervised learning, the network receives inputs at the bottom level and a category input (supervision) at the top level, and it learns which causes in the world belong to each category. If you train a network with unsupervised learning, a network only receives bottom-level inputs, and it devises categories of its own at the top level. But even during supervised learning, all nodes of a network except the top node still learn in an unsupervised fashion. HTM networks support both types of learning.

    Q:

    How can I be sure that my HTM Network is even using time?

    A:

    The HTM model assumes a world where underlying causes change slowly with time. Thus the HTM algorithm always attempts to find and exploit temporal correlation. It may be that your problem does not even require this. In this case the algorithm may still work and give reasonable results. One definitive test is to set the max group size parameter to 1 for all the nodes and re-run the learning. If performance is not affected, or improved, then your problem is not helped by the algorithm's temporal assumptions.

    Q:

    When working with NuPIC, my STDOUT.TXT file is growing very large. What is causing this?

    A:

    When debugging your own network, you may have added a few print statements into the code you were debugging. Check your code, including any custom PyNodes, for any print statements, as this is the most common way to increase the size of the STDOUT.TXT file.

    Q:

    I'm getting 0% correct on all of my inferences. What could I be doing wrong?

    A:

    Something very basic is wrong. Please see the Programmer's Guide or the Debugging section for more information.

    Q:

    My network is coming up with too many coincidences. What should I do?

    A:

    Increasing the maxDistance parameter reduces the overall number of coincidences at Level 1 nodes, which in turn may reduce the number of coincidences at higher levels. However many applications really do require a large number of coincidences. The Pictures task generates over 40,000 coincidences per node at level 2 - this is not necessarily wrong. In these cases increasing maxDistance too much will have a negative impact on performance.

    Q:

    My network is coming up with too few coincidences. What should I do?

    A:

    Try decreasing the maxDistance parameter of the bottom-level nodes. Doing so tells the nodes to consider similar inputs to be distinct, which increases the number of coincidences at all levels of the hierarchy.

    Q:

    All of my coincidences are in one group. What should I do?

    A:

    Try decreasing the topNeighbors parameter at the affected nodes. This causes the nodes to form smaller groups, leading to more groups overall.

    Q:

    I have so many groups, many of which contain only a single coincidence. What should I do?

    A:

    Try increasing the topNeighbors parameter. Usually, this leads to larger groups and fewer groups overall. Remember that small groups with infrequently-encountered coincidences will almost never be highly-ranked during inference, so having many singleton groups is not necessarily a problem.