The notion that the more data, the slower the system – ain’t always true.
My favorite way to explain this very important phenomenon involves the familiar process of assembling a jigsaw puzzle.
The first piece you take out of the box and place on the work surface requires very little computational effort. The second and third pieces require almost equally insignificant mental effort. Then as the number of pieces on the table grows the effort to determine where the next piece goes increases as well. But there is a tipping point where the effort to determine where to place the next piece gets easier and easier … despite the fact the number of puzzle pieces on the table continues to grow.
Well isn’t it interesting, although obvious, that those last few puzzle pieces take nearly as little effort as the first few!
I have witnessed this.
This has a slew of ramifications.
This does not apply to all domains. This behavior requires: (a) observations from the same universe; (b) observations with enough features to enable contextualization; (c) observations in which these features can be extracted, enhanced and classified; (d) sufficient saturation of the observational space; and (e) enough smarts to stitch these puzzle pieces together.
Context accumulating systems, fed appropriate observations, can be expected to have this behavior.