According to Brynjolfsson and Mcafee (2017), machine learning (ML) is the most important general-purpose technology of our era, similar to previous general purpose technologies such as steam power, electricity and the combution engine.
They argue that important general-purpose technologies, such as ML, have a transformational impact on the economy and business, both directly, but also indirectly by enabling complementary innovations and opportunities.
There are two reasons why ML is such a big deal:
Humans cannot automate a number of tasks since they cannot explain how to carry out these tasks. In such cases, traditional software cannot be developed, since it requires the software developer to understand the tasks at hand in order to code the predefined sequence of instructions.
Ready Polanyl: Humans know more than they can tell.
On the other hand, ML systems are excellent learners, so with ML we can automate these tasks even if humans do not know the sequence of instructions. Why? Since ML can learn from data.
Brynjolfsson and Mcafee (2017), define Machine Learning (ML) as “the machine’s ability to keep improving its performance without humans having to explain exactly how to accomplish all the task it’s given”.
ML comes in two major categories:
Supervised learning systems are trained using “labelled” training data in order to categorize new, unlabeled data, based on the training.
More precisely, a supervised learning system use a large set of training data showing both the input (for example, pictures of dogs and cats) and the output.
After the system has been trained with the training data, the system should be able to do the mapping between input and output itself. That is, given new input (new pictures of dogs and cats), the system should be able to predict the correct output (label of dog or label of cat).
A somewhat simplified example of how a supervised learning system address a given problem:
First, a ML algorithm and set of training data need to be selected than are suitable for the problem to be addressed.
The selected ML algorithm could look like this: $y = kx+l$ (but represented as code in order to be run in a computer). The algorithm in this case is called “linear regression”.
The training data could consist of one input ($x$) and one output ($y$) value, in the form of a value pair ($x,y$), for example, $(1,3), (0,2), (-1,0)$, etc.