Why is renormalization necessary in finite theories? Michael Brown made the following comment here:

The modern understanding of renormalization (due to Kadanoff, Wilson and others) is hardly controversial and has nothing really to do with infinities. It is necessary even in completely finite theories, but the fact that it fixes infinities in QFT is a bonus.

Can anyone explain what this means?
 A: Here's one way of interpreting this statement (this is essentially an elaboration of the comment by use Learning is a mess).  The basic idea behind Wilsonian renormalization is that renormalization can be regarded as the process by which we understand how the theory effectively behaves at different scales, like momentum scales.  Disclaimer: the math in this answer is schematic.
Consider a quantum theory of fields $\phi$ with a hard momentum space cutoff $\Lambda$.  Such a QFT would be described, in the Euclideanized functional integral picture, by its partition function
\begin{align}
  Z(\mathbf u, \Lambda) = \int [d\tilde\phi]_0^\Lambda e^{-S[\mathbf u, \Lambda,\tilde\phi]}
\end{align}
Here, $\mathbf u = (u_1, \dots, u_n)$ represents the parameters on which the action of the theory depends (like coupling constants), $S[\mathbf u, \Lambda,\phi]$ is the Euclidean action, and
\begin{align}
  [d\tilde\phi]_{k_a}^{k_b} = \prod_{k_a<|k|<k_b}d\tilde\phi(k)
\end{align}
schematically denotes integration over field degrees of freedom corresponding to momenta between $k_a$ and $k_b$.
Now, let any real number $0<s\leq 1$ be given.  We will call $s$ the "scale."  Then we note that we can split the functional integration measure into an integration over momenta in the range $(0,s\Lambda)$ and those in the range $(s\Lambda, \Lambda)$, and we imagine defining a rescaled action $S_s[\mathbf u, \Lambda,\phi]$ by
\begin{align}
  e^{-S_s[\mathbf u, \Lambda, \phi]} = \int [d\tilde\phi]_{s\Lambda}^\Lambda e^{-S[\mathbf u, \Lambda,\tilde\phi]}
\end{align}
We can now write the original partition function as an integral over only the momentum modes in the range $(0, s\Lambda)$ provided we use the action $S_s$;
\begin{align}
  Z(\mathbf u, \Lambda) = \int [d\tilde\phi]_0^{s\Lambda} e^{-S_s[\mathbf u, \Lambda,\tilde\phi]}
\end{align}
and we say that we have "integrated out" the higher momentum modes.  The action $S_s$ can now be regarded as the action that effectively governs the physics for momentum scales below $s\Lambda$.  
Let's call $S_s$ the action "at scale $s$".  Then in some situations, the action at scale $s$ can simply be related to the original action at scale $s=1$ by taking the couplings to depend on the scale $s$;
\begin{align}
  S_s[\mathbf u, \Lambda, \phi] = S[\mathbf u_s, \Lambda, \phi_s]
\end{align}
This is often referred to as the "running of the couplings."  In other words, the process of renormalization which leads to the running of the couplings is simply the process by which we examine how the field theory effectively behaves at different scales; this is conceptually distinct from the issue of removing infinities.  
