Say you want to write a compiler whose input is "bound" to a source file; when the input changes, you don't want to recompile the entire file all over again. Instead, you memoize at AST node boundaries, trace dependencies between AST nodes, and then "clean" nodes when they become dirty, detecting changes that affect other nodes to clean them also.
To be honest, I find these data binding frameworks to be very boring, since we've been doing this forever (I did my dissertation on one myself). We really should be looking at the next level that supports incremental processing of more complicated updates.
There is a lot of research in this area, concurrent revisions at MSR or self-adjusting computations at CMU to name a couple. Also, if you've seen Bret Victor's videos and wonder how any of those could be implemented for real, then this is where we need to go. It will be a fun trip though.
Hard to explain concisely but it allows fairly complex updates to be expressed very efficiently - see this animated chart for example. The path is created using an expression, which is reevaluated during animation and whenever the dimensions of the chart change: http://www.ractivejs.org/examples/animated-chart/