Data Fitting

Modern artificial intelligence rely on data fitting. For some, the approach seems shallow, for the result is a description that summarizes data, rather than a mechanism that generates them. The success of modern artificial intelligence requires a justification of data fitting in order to explain how a seemingly shallow method turns out to be highly efficient. One doesn't want to invoke the analogy of catching a baseball. Even if a baseball player has no idea of classical mechanics, the ball can be traced by following its trajectory. Data fitting is deeper than this.

First it must be emphasized that data fitting is a essential tool even for hard sciences. Planck had no idea how to generate light quanta, but based on curve fitting derived with the quantum hypothesis, he successfully launched the quantum revolution. The beauty of his approach is the small number of parameters that can explain a host of radiation patterns. Although data fitting can not explain fundamental mechanisms, it can indicate the correct direction by ruling out inconsistent hypotheses.

But purists contested. They say that once we have physical laws, data fitting is no longer necessary. Quite the opposite, since physical laws don't provide specification for configuration, even if all physical laws are known, we still can not pin down whether Biden or Trump won the 2020 election without electoral data. That is, physical laws provide restrictions on possible universes, but don't supply the answer to which. A way to elucidate this is to observe that both the hypotheses that Biden or Trump won the 2020 election are consistent with physical laws.

Since configuration can not be derived from physical laws, and there is arbitrariness in the formation, data fitting is a economical way to describe observed configuration. We can know a face without detailing all bio-molecules. Further, because there may be regularities in data, prediction can be made without cumbersome integration of microscopic mechanisms. Data fitting is a efficient method to cut through complexities.

Hope this provides some justification for data fitting so that it may no longer be considered shallow, but necessary.