Great work. It is a nice accomplishment and it is amazing that an end application is able to do this on top of the infinite flexibility which is the foundation for it. We have full access to the foundation. I've posted a video response and am interested in problems which HTMs may be adapted to. Thanks Geometry is beautiful!
1 don't you think you need to exclude the background black box so, that it will not recognize the objects as cube? just using the solid black or white background..
1) send me the images and I will make them all the exact same size with black for the entire background
2) there is enough with HTMs that your empty bored modalities should be packed full - busy thinking - the music helps you stay in this abstract landscape.
1. I actually observed (and on account of my imagining the nature of its perspective) that the edges were being included in the HTM's intake of the stimuli, but I thought it was too tedious a job to fix (with little gains). If you're a willing partner in my crime, then I won't stop you! You should receive an e-mail tomorrow, hopefully.
2. This choice of song actually originated in considering the nature of the video -- nothing too engaging, but rather contemplative.
Thanks for the comment. The brunt of the labor involved collecting the images which was extremely tedious work.
In time, after I've established another data source, I will add wire-frame style images to the data set and see what kind of results are had whilst testing "real world" images, as naturally imperfect as they are.
Incidentally, I have already managed to attain 100% accuracy by merely adding more polyhedraorg images of dodecahedra and icosahedra (the more complex forms) for training.
Images of cows, toy ducks, and so forth would not be pertinent to testing whether my HTM would be able to identify any one of the regular polyhedra.
What I mean by "'real world' images" is physical embodiments of the geometric forms in the world, photographed and digitized so that they would be suited to testing on my HTM. That is what my meaning, so contextualized, was.
well, if i get you correctly, i assume it will work even better with this because of its principles of work. but any tests and publishing the results are appreciated!
That is actually an interesting point that I have considered. Would it not be best to use images of "really existing" regular polyhedra? I thought. And then I answered: but if all I am giving to the HTM is the polyhedra themselves, then resorting to messy, imperfect data like that granted by the "real world" would only result in imperfect results. Let's face it: in reality there is no actually existing Platonic form. So with that I decided upon using the idealizations themselves instead.
@argumzio good recognition software should still yield good results with messy or incomplete inputs. if it can only recognize an ideal and perfect completely typical presentation then it isn't very useful. fortunately, one of the greatest achievement of the HTM design is that it is still very reliable under less than ideal conditions (just like a brain).
@RRRRussia Yes, this is one of the hallmarks of HTMs as they are now and certainly will continue to be. I chose the quasi-ideal inputs because of their accessibility. I don't have time to amass a collection of messy un-Platonic inputs for Platonic solids in order to teach an ignorant network what they are. However, pure recognition isn't enough for me. I hope that one day we can teach HTMs something like this (2D inputs) and that it will be able to output a 3D (and higher) construct, like we do.
Great work. It is a nice accomplishment and it is amazing that an end application is able to do this on top of the infinite flexibility which is the foundation for it. We have full access to the foundation. I've posted a video response and am interested in problems which HTMs may be adapted to. Thanks Geometry is beautiful!
Digikong 2 years ago
ok, the good one.
but:
1 don't you think you need to exclude the background black box so, that it will not recognize the objects as cube? just using the solid black or white background..
2 boring music
thank you.
Sergiuss555 2 years ago
Thanks.
1. The files are in .png format. The toolkit behaves a little strangely -- seemingly filling in the empty space with black.
2. My object was to bore you, not to entertain.
You're welcome.
argumzio 2 years ago
1) send me the images and I will make them all the exact same size with black for the entire background
2) there is enough with HTMs that your empty bored modalities should be packed full - busy thinking - the music helps you stay in this abstract landscape.
Digikong 2 years ago
1. I actually observed (and on account of my imagining the nature of its perspective) that the edges were being included in the HTM's intake of the stimuli, but I thought it was too tedious a job to fix (with little gains). If you're a willing partner in my crime, then I won't stop you! You should receive an e-mail tomorrow, hopefully.
2. This choice of song actually originated in considering the nature of the video -- nothing too engaging, but rather contemplative.
Thank you for your comments.
argumzio 2 years ago
Off-topic: I find the music very engaging. :)
zalmoxe 2 years ago
wow! good job
amystarr311 2 years ago
Thanks for the comment. The brunt of the labor involved collecting the images which was extremely tedious work.
In time, after I've established another data source, I will add wire-frame style images to the data set and see what kind of results are had whilst testing "real world" images, as naturally imperfect as they are.
Incidentally, I have already managed to attain 100% accuracy by merely adding more polyhedraorg images of dodecahedra and icosahedra (the more complex forms) for training.
argumzio 2 years ago
Comment removed
Sergiuss555 2 years ago
..the deleted comment was posted as a result of my brain cpu hanging ..
what about the standard set of cow, toy duck, etc? they look pretty real-world, uh?
Sergiuss555 2 years ago
Images of cows, toy ducks, and so forth would not be pertinent to testing whether my HTM would be able to identify any one of the regular polyhedra.
What I mean by "'real world' images" is physical embodiments of the geometric forms in the world, photographed and digitized so that they would be suited to testing on my HTM. That is what my meaning, so contextualized, was.
Cheers.
argumzio 2 years ago
well, if i get you correctly, i assume it will work even better with this because of its principles of work. but any tests and publishing the results are appreciated!
Sergiuss555 2 years ago
That is actually an interesting point that I have considered. Would it not be best to use images of "really existing" regular polyhedra? I thought. And then I answered: but if all I am giving to the HTM is the polyhedra themselves, then resorting to messy, imperfect data like that granted by the "real world" would only result in imperfect results. Let's face it: in reality there is no actually existing Platonic form. So with that I decided upon using the idealizations themselves instead.
Thanks!
argumzio 2 years ago
@argumzio good recognition software should still yield good results with messy or incomplete inputs. if it can only recognize an ideal and perfect completely typical presentation then it isn't very useful. fortunately, one of the greatest achievement of the HTM design is that it is still very reliable under less than ideal conditions (just like a brain).
RRRRussia 1 year ago
@RRRRussia Yes, this is one of the hallmarks of HTMs as they are now and certainly will continue to be. I chose the quasi-ideal inputs because of their accessibility. I don't have time to amass a collection of messy un-Platonic inputs for Platonic solids in order to teach an ignorant network what they are. However, pure recognition isn't enough for me. I hope that one day we can teach HTMs something like this (2D inputs) and that it will be able to output a 3D (and higher) construct, like we do.
argumzio 1 year ago