Hello, I have implemented HTM in Swift for iOS/MacOS and am currently playing around with creating a test bench for simulations to get the best performance out of the system. The project will be finalised in a couple of weeks.
I have moved in a slightly different direction by implementing a own version that I think makes more sense, but maybe someone would be interested in the plain HTM binaries? Configuration is made by a XML file that really requires no programming skills. Performance is pretty good at 13.9MB per region of 2048 cells.
Here is a screen from Alpha and the GUI will need some additional work before being usable (Excuse the black bar, but there was system information that is irrelevant for anyone but me).
Also excuse the rather nasty colors, they are there so I can see changes when the simulation runs really fast. Also my deepest apoligies for using a “windows like” blue in the background will actually change that to a more Mac-isch silver/gray gradient right now.
Green = Predicted
Purple = Active
Orange = Inactive
If anyone would be interested in binaries I can add some more stuff to it with encoder selection and stuff like that.
Yes, I am also very interested. I also like your GUI implementation, despite the colors. Can you make the codebase also available? Without the codebase as open source, it will not be possible for us to continue making improvements and modifications being discussed in the HTM-Theory threads.
I can make the codebase open source, just have to remove some stuff from it like the way I form predictions and stuff like that. Will update this thread once the project is on GitHub in some weeks.
The original idea was to have the HTM as a box (Bin. download) where you can select your encoder and then input a file or stream of data using the standard set of encoders.
Just a small update to give a flavour of how things could look in the end.
I have still not come up with a understandable way to link the input graphically, I do enjoy working on this though so I will see it through to the end.
If anyone want a mac HTM ‘app’ functioning like a blackbox for their data, please do give suggestions on functions/features you would like to see.
In the top right corner I want to add a real time graph, similar to what Matt uses in the HTM school videos to graph the incoming data and show predictions/activation patterns in real time.
Refined the GUI a bit today. Have still not figured out how to show inputs (other than maybe graph form) or how to show multi-layer regions as space is limited.
I don’t have a boss so I feel this is meditative to post progress here, so I feel accountable to myself while maybe someone will find it useful and offer comments/input. If you think this is totally uninteresting just ignore the topic
The “dark bar” at the top will be a graph over time, showing prediction success rates and other data points.
Right now the input is a hardcoded interface, I am thinking about how to solve this graphically, but nothing clean has appeared to me yet.
So this is how I will illustrate the regions… I am working on a algorithm to create a visual representation from the region by subsampling and “making it look good” but with a subset of “real” data.
Not sure yet if to use 2D or 3D visuals.
Reduced loading time from 36minutes to… 86 seconds
I know one should not prematurely optimise things, but it kinda sucked waiting half an hour for 1!! region to load.
My plan was to remove stuff, but instead the SP and TM became one and lives in the region. I will probably keep this as a separate project maybe and create a vanilla HTM behind the GUI and release it as open source.
Just wanted to say sorry for lack of updates. I am currently in the process of relocating to Vienna, Austria so expect further updates in a week or two. I think I have solved the input visualisation issue in an elegant manner though.
(Excuse the black bar, but there was system information that is irrelevant for anyone but me).
will actually change that to a more Mac-isch silver/gray gradient right now.
