Getting started, looking for tips

Hi everyone! I recently heard about HTM and I’m really hyped about it. To facilitate my learning, I want to program a complete (but mini) version on my own and test it on a sample problem. I have tons of questions! Any advice or discussion you can offer is really appreciated.

  1. What is a good dataset to start with to test basic SP and TM functionality on? I have some familiarity with the MNIST dataset so that’s what I default to usually, but I wonder if that may be too complex for a beginner.

  2. I went through the HTM school videos so I know the basic architecture of HTM, but I wonder what are the typical values for the various numeric parameters in the model? These were of course not given in the videos. I.E. permanence update rate, overlap thresholds, etc.

  3. Practical issues: I’m not sure a priori how CPU-intensive the HTM learning process will be–will a typical home laptop be capable of even running it?

Thanks for reading this far! If you can help me out, thanks again for that as well! I will probably reward your kindness with additional questions :wink:

2 Likes

It seems like some of your questions might be answered in Read this first. In particular, there is a link to this example with data: Algorithm API notebook for new-ish interested people.

6 Likes

I’d recommend looking at the famous HotGym example. It uses streams of energy use values from a gym to do anomaly detection, with another instance for forecasting.

Here’s the code base:

Here’s another great video walkthrough by @rhyolight:

2 Likes