about how they schedule their initial step sizes.  In particular, they will try
to take step sizes that are similar to previously used sizes via a trust region
style scheduling.

backtracking_line_search() by making it use 3rd degree polynomial interpolation
after the first step.  Also made it more robust to alpha inputs with improper
signs.

computations done in f() when computing der().

tests so they work on older versions of visual studio.

that allow you to easily set uniform upper and lower bounds.

functions with the optimization code and forced the use of function pointers.
This was to avoid triggering a bug in gcc 4.0.  Since that compiler is no
longer officially supported by dlib I've removed these checks to increase
usability.

bounds.

false alarming vs getting a correct detection.

is missing the HTTP/1.1 version field.

errors made by an approximate separation oracle.  In particular, the solver
will now check the output of the separation oracle against the cache and if the
cache gives a better value, even when we would otherwise not use the cache, the
cache value is used.  Similarly, we can output the truth psi vector to avoid
outputting a psi with a negative risk.  All this stuff only happens when the
cache is enabled, if its disabled then the outputs of the separation oracle are
used without any kind of modification.

which detection template generated a detection slightly more accurate.