Q&A with Carey 1- How do you know what concepts each "bottom layer node" corresponds to? You turn it on and let the system vibrate, and the top layer will (tend to) represent it. 2- If you needed an algorithm to learn a specific concept, how would you know which layer of the DBN corresponds to the level of abstraction of that concept? It's empirical at the moment - trial and error, intuition. It's still early stages. 3- If you needed an algorithm to learn a complex concept AS OPPOSED TO another, for example given a large set of still pictures, identify safe animals that don't bite versus dangerous animals that might: would you train it on a set containing both types of animals? Or would you train it on the safe ones only? It depends if you have large labelled datasets. If you don't, you could train it on a set containing both, and couple the results with a classifier algorithm such as decision tree. You would still need some form of labelling for that.
![Baby Kia - GLOVES & DRACS (BLACK AF1s) [Official Music Video]](http://i.ytimg.com/vi/ajm8LeB0qQk/mqdefault.jpg)
Best explanation on Ising model and Boltzmann machine online!
This is the best introductory video I've seen on this topic, now to figure out how the convolution fits in...
Q&A with Carey
1- How do you know what concepts each "bottom layer node" corresponds to?
You turn it on and let the system vibrate, and the top layer will (tend to) represent it.
2- If you needed an algorithm to learn a specific concept, how would you know which layer of the DBN corresponds to the level of abstraction of that concept?
It's empirical at the moment - trial and error, intuition. It's still early stages.
3- If you needed an algorithm to learn a complex concept AS OPPOSED TO another, for example given a large set of still pictures, identify safe animals that don't bite versus dangerous animals that might: would you train it on a set containing both types of animals? Or would you train it on the safe ones only?
It depends if you have large labelled datasets. If you don't, you could train it on a set containing both, and couple the results with a classifier algorithm such as decision tree. You would still need some form of labelling for that.
Pure joy ride! Multiple Aaha moments throughout the talk. The slides would have taken ages to create given the bulk of animations in them ...
yeah, i also wonder what kind of software tools he uses to produce these slides