I am researching anomaly detection in electromagnetic emissions from IoT devices and considering using HTM or principles from the TBT for this purpose.
I have a few key questions:
HTM vs. TBT:
What are the fundamental differences between HTM and TBT?
Is HTM a subset of TBT, or does TBT introduce new computational principles?
Are there any practical implementations of TBT, or is most work still based on HTM?
Anomaly Detection in EM Signals:
Can HTM/TBT be used to detect anomalies in continuous EM signals from IoT devices?
Given that EM signals have spatiotemporal complexity (shifting frequencies, noise, weak signals), how well does HTM handle unsupervised anomaly detection in such data?
Existing Examples & Research:
Are there previous projects where HTM was applied to RF signals, spectrum analysis, or similar time-series data?
Do you have code samples, research papers, or recommended resources for such an application?
I’d greatly appreciate any insights, references, or recommendations!
Thanks for your time and for your work.
Welcome aboard! It has been a while since you posted your questions, so I’m going to jump in with some extremely unofficial answers. So, YMMV…
The Thousand Brains Project (TBP) is the cutting edge of Numenta’s TBT work, so I’ll mostly use that as my focus below. It’s very early days for the TBP, so a lot of issues (both theoretical and practical) are still pretty open.
However, the TBP is following the TBT’s general themes, as laid out in A Thousand Brains (just as HTM is described in On Intelligence). There are also a lot of TBP videos which provide both introductory and overview information, cover both historic and current brainstorming sessions, etc.
In brief, however, the TBP is mostly concerned with simulating networks of cortical columns. So, there will be large numbers of Actors (i.e., light-weight processes, communicating with each other via message passing).
At this point, most of the work has to do with developing and exercising Learning Modules, using simulated Sensor Modules in a test framework. However, work on “real” Sensor and Effector Modules will soon be ramping up, in part because of an upcoming robotics hackathon.
Clearly, some of these modules will have to deal with temporal sequences. However, the current focus is on using sensorimotor-based interaction to explore, model, and recognize physical objects (e.g., coffee cups, staplers) and behaviors.
To support this, the network will include both hierarchical and heterarchical connections. The Learning Modules will also need to perform exploration planning, etc. In short, HTM has mostly been subsumed into the broader TBP architecture.
I don’t know of any current production use of HTM. Numenta used to have a product called HTM Studio, but the web page says “HTM is part of our legacy research and is no longer maintained by Numenta”.
This doesn’t sound like anything I’ve seen the TBP folks discuss, but here are some partly baked ideas and questions (no extra charge :-}):
You haven’t said what kind(s) of anomalies you have in mind. Are we talking about drop-outs, excessive traffic, noise spikes, or what? In any case, I could imagine creating a Sensor Module that would be hard-coded to convert EM signals into streams of event and/or statistical data. A set of Learning Modules might then be able to explore the data, model and recognize anomalies, etc.
@Rich_Morin great response! Thanks for jumping in here and answering @yahavka12 questions
I don’t have a lot more to add to this, just a few more thoughts and links:
Rich is correct in that the TBP implementation is currently focused on modeling static objects and learning and recognizing them by moving a sensor. We are planning to incorporate a time component into our models but this is work in progress (some more thoughts on this in @nleadholm post here Extending and Generalizing TBT Learning Models - #27 by nleadholm ).
If your application is focused on anomaly detection in time series data, not sensorimotor learning, then HTM is currently the better approach to look into. There is a separate HTM forum here. Numenta recently changed the license of the associated code to an MIT license but the code is not being actively maintained.
If you want to read more about the TBP approach, you can check out our whitepaper. For more info on HTM and anomaly detection work from Numenta (The TBP is a non-profit separate from Numenta now), you can have a look at Numenta’s papers.
