Analytics SolversA quick look at Google Trends searches for "artificial intelligence", in Argentina and worldwide, confirms that the topic is part of many everyday conversations.
The world consistently searches these terms more than Argentina, and the gap widens for "machine learning", a concept often credited with almost magical powers to solve business problems.
Artificial intelligence (AI) seeks to make machines imitate humans, and Machine Learning (ML) is one of its key dimensions. That is why it is important to understand what ML does and does not guarantee when used to solve business problems. There are four classic approaches to defining AI: a machine can act and think humanly and rationally.
One of the basic capabilities such a machine needs in order to act and think is to learn. To imitate humans, it must first interact with its environment and, from it, detect and extrapolate patterns in the data it receives. Hence the term machine learning.
What does learning mean in these terms? Every observable aspect of reality is generated by a series of instantaneous or delayed causes that accumulate and amplify through thresholds in countless ways. Observable data change state when all or some of those causes set things in motion. If we could understand what, when, how and why the aspects we care about change, we would be able to influence events at will.
In mathematical terms, this means that all the explanatory variables causing an observable event, which we will call the target, are related to it through an unknown function.
Supervised Machine Learning only tries to find that function using the variables it is given. This leads to three very common situations.
1. Garbage data
If the data are garbage, the output is garbage, even with the most complex and powerful algorithm available. In this case, machine learning cannot learn anything.
2. Fully specified target
If the target is fully specified (all explanatory variables are included), machine learning could learn it with minimal error, despite the enormous complexity that may involve.
3. Partially specified target
This is the most common situation: only some relevant variables can be included, proxies for the ones needed, or redundant variables that interact heavily with others that have no clear relationship with the target.
Machine learning only guarantees removing bias and minimizing error, and the latter is what people usually misunderstand. That error is the theoretical error achievable when the variables given to the algorithm try to learn the target. This theoretical minimum does not depend on Machine Learning but on the variables: ML only guarantees getting close to it, and it may still be unacceptable. Bias, in turn, represents the memory capacity of the function when trained. Memorizing means learning non-generalizable details; that is how the well-known overfitting happens, and machine learning fails.
In short, given a set of explanatory variables and a target, Machine Learning only guarantees that it will try to learn the target with the lowest possible error and generalize that learning to future data. It does not promise to learn the target perfectly, even with highly complex models: it will learn a lot, some or little, depending on what the included variables allow. It does not perform magic or find patterns that do not exist in the data.
Machine learning is a laborious, interdisciplinary process. If a problem was solved just by loading a dataset and running an algorithm, the problem was trivial. And in general, that is not the case.