June 10, 2009

Alright, so last night we narrowed it down to four ideas for a project:

1. Multi-modal self-association
2. Multi-modal object recognition
3. Self-discover using an HTM
4. Object tracking using an HTM

The real debate between the ideas has come down to one issue: to use an HTM or not to use an HTM, that is the question. It is my strong opinion that, as much as I'd like to, an HTM is way too ambitious for us to accomplish in one summer. It is my opinion that we choose one of the first two ideas listed and save the last ones for future work. We have been and will continue to discuss this today, and I hope that by the end we can settle on one or two ideas. I think that I have successfully swayed Taylor and Ugonna to my point of view. Jivko is currently collecting papers that apply to the first two topics and will be sending them to us tonight. Him and the other two grad students (Matt Miller is not an advisor but has experience with HTM's and works in the lab) share my opinion. Taylor seems disappointed, but I think he agrees with my logic that if we try to build an HTM in a summer, we will end up scaling it back so much that it will not be what we originally intended and, more importantly, not publishable. That is just about all for now, hopefully tomorrow I can say for sure what our topic will be.

ConnorSchenck - 12 Jun 2009 - 13:03

An HTM is a hierarchical temporal memory. Essentially it is a group of nodes arranged in a pyramid fashion. The nodes at the bottom take in the raw input data and pass data to the nodes aove them etc. Every node operates in the same manner: they first classify the patterns they see spatially and then classify them temporally and pass that up to the node above them. A good example of how this works is if you imagine that you are a node. You get different pieces of colored paper from the nodes below you. None of them are the exact same color as the others. But you can still classify them. You put them in buckets labeled "red" "green" "yellow" etc, and none of the pieces of paper in the red bucket are the exact same shade of red, but they are all red. After that you start to notice patterns in the order of colors you get. You might notice, for example, RRGY for red red green yellow. You notice that red is usually followed by another red, which is followed by green, and finally followed by yellow. So, over time, you start to notice many patterns such as this. Now, when you create something to hand to the node above you, you hand up a piece of paper saying the name of the pattern such as "RRGY" and the node above does the same thing with that that you just did.. That is the essence of an HTM. There is also a prediction factor, but I won't get into that now.

ZayiraJordan - 11 Jun 2009 - 15:24

yeah, what is an HTM?

MinaChoi - 10 Jun 2009 - 16:01

what is HTM?