approximate pseudo inverse solver (see: http://en.wikipedia.org/wiki/Moore�Penrose_pseudoinverse ) solves ( t(a) a + epsilon I ) x = t(a) y (not if a

statistics accumulator used by BTable builds stats on categorical variables with a very large number of levels @author johnmount

very low dimensional re-encoding of levels @author johnmount

only applicable to symmetric positive definite matrices (or their negations) @author johnmount

@author johnmount

data carrier, weaker than a sparse vector in that you can not jump to an arbitrary coordinate @author johnmount

for testing, force calls of the merging code to test intermediate paths. Deliberately inefficient @author johnmount

for testing, force both reducers into serial op @author johnmount @param <T> @param <S> @param <Z>

not as good as general conjugate gradient @author johnmount

The standard way to encode a categorical variable- build one indicator function per value (we don't supress one level). @author johnmount

@author johnmount

Interface for fast tightly coupled summation across fixed x,wantGrad,wantHessian varying ExampleRow into an accumulating VEval. Designed usage: summin

some hard-coded paths, so can not run two of these simultaneously @author johnmount

Newton step optimizer (set up to maximize) @author johnmount

return -x.x (maximum value 0 achieved at x = 0). @author johnmount

For classes implementing this interface the following are true: 1) a.observe(T t1,t2, ...,tk) is equivalent to a.observe(T ta,tb,...,tz) where a,b,

keep a uniform sample of items seen @author johnmount @param <T>

caches recent evals (de-bounce fn evals so other code can be written simply) keeps best eval (maximal) @author johnmount

Sigmoid loss function (maximizing this yields logistic regression) for multinomial classifier (all classes have variables) f(x) = sum_{data pos} log(

Class for slow loosely coupled summation of VectorFns. designed usage: summing two unrelated functions (like data score and regularization penalty) @a

@author johnmount @param <T> type being observed @param <Z> type of observer that will work in parallel @param <S> type of observer that will pre-sca

Result of a VectorFn eval or the accumulation structure for a LinearContribution @author johnmount