Hello! I need some advice on how to set up a model that can predict the trajectory of a projectile using a hierarchy to establish a sort of generalization process. The result should be some kind of an extremely extremely rudimentary conception of gravitation (simulated, of course). Before asking specific questions, Some Background:
The idea was inspired by Michael Ferrier's bouncing ball demo posted on youtube (if you haven't seen it: http://www.youtube.com/watch?v=YeBC9eew3Lg). I thought this video demonstrated nicely a problem domain that I think HTM CLA technology should be well-suited for, which is spatial-temporal reasoning. It was awesome how the model predicted how the ball should 'fall' down the retina, but I noticed that the prediction would only work if the ball starts in the same spot every time, and that there is no generalized understanding that the object moves down the retina. If the ball is moved to the right, upper corner of the retina, the activated neurons would be incapable of accurately predicting that the ball will fall because those neurons have no experience with the temporal happenings of a ball in the air. My thought is that this could be improved upon if a network is established to include independent regions handling each physical aspect of the ball. In the case of a projectile with both x and y axis motion, for example, you could have a region for handling the changing x-axis position, a region for handling the changing y-axis position, and maybe a region for handling velocity and/or angle and/or other important dimensions of the event. Then, each region is responsible for handling a component of the event, so the output from each region could be put into a single region which maps the combined output of each region to a single state (position of the ball), which is the prediction of where the ball is in the next time step. I think there would have to be some kind of supervision process, though, to get the final region to map combinations of inputs to the correct prediction efficiently. Is the temporal pooling routine sufficient? I think this would be an interesting project because it demonstrates/forces us to think about how visual information is divided by the visual areas of the brain and conquered by higher regions. It also demonstrates generalization processes. If it works, the network wouldn't have to experience nearly every possible happening like a single region would. Each region just needs a sufficient amount of experience to be able to get the whole network to be able to make accurate predictions of projectile motion. I think.. Anyway, questions: 1) Are there any ideas on how I could split the initial input in a brain-like way so that the appropriate components of the input can be pushed into their respective regions? 2) How does one connect regions? 3) If I want to provide the input in the form of a matrix with 0's (1's representing the position of the ball), how would I go about that? Thanks for reading all of this! Any help/conversation would be great. -Josh Rose
_______________________________________________ nupic mailing list [email protected] http://lists.numenta.org/mailman/listinfo/nupic_lists.numenta.org
