... system.1.1
It is not essential that the input be perceptual. However, perceptual modalities are rich, expressive, intuitive and non-obtrusive. One could take other measurements if they help infer behaviour, internal state or intentionality.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... game)2.1
Of course, the individuals need not be in the same physical space and could be interacting through a virtual environment.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... integral5.1
Either analytically or by sampling techniques
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... space5.2
This is the typical task of unsupervised learning.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... output5.3
This is the typical task of supervised learning.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... density.5.4
A proof by example can be found in the appendix.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... optimization.6.1
We shall focus on maximization in this document but most arguments should apply to minimization problems as well.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...$\Sigma$6.2
In this sense, $\Sigma$ acts like a variational parameter.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...>06.3
Here, a is also only an integer traditionally but this is not necessary for the following derivation. Also note that a does not truly have to be positive since $\sin(-x) = sin(x+\pi)$ can be used to factor out the negativity.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... covariances7.1
In fact, step 1) and step 2) can be performed simultaneously since they only involves taking derivatives and setting to 0.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
...[5] 7.2
A classic example of this is that of dystal learning to recover the angle at which to throw a ball when trying shoot it a pre-specified distance. Since throwing a ball at 90 or 0 degrees is equivalent (and any other (90-x) or (x) combination), the expected value is to always throw the ball at 45 degrees which is incorrect and would have probability $\approx 0$.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
... constraints11.1
In fact, in assuming a joint density, the model is a parent of the ${\cal X}$ data instead of being a co-parent with it.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Tony Jebara
1999-09-15