If v1 is off, does the existence of h1 affect v2? 

A: not a lot. h1 will want to turn off.













what if the bias were -6? what if it were +6?

A:

-6: h1 will almost always turn off so it has almost no effect.

+6: h1 is almost always on, so its connections are almost like biases.














A Restricted Boltzmann Machine is often said to "reconstruct" its data. That sounds as if an RBM were a deterministic feedforward network with input units (for the data) and output units (for that reconstruction). However, it isn't. Why is this "reconstructing" a reasonable concept anyway? 

A: We sample a state for the hidden units, given the visible units being set to the training case. Then we sample a new state of the visible units. Usually, this will be similar to the original, hence "reconstruction".








Could we construct a deterministic feedforward network that does have such units, and that performs approximately the same reconstruction as a given RBM?

A: yes, we did this on the blackboard. It has logistic units.







Questions that we didn't get to:

find the conditional probability for a gaussian visible rbm.

video 12d, 4:23: how does that unit represent where the top of the 2 is? Elaborate. Remember that the unit takes on stochastic binary states.

video 12d, 5:57: why is that obviously for 8s?