Historically we have been told that computers need to be programmed in order to perform specific tasks. However, in recent years, scientists have developed algorithms that make possible for computers to act, based on patterns and statistical trends instead of commands programmed by humans. In short, Machine learning algorithms use historical datasets to make predictions and decisions without being programmed for them. This technology is particularly useful for applications where it wouldn’t be feasible to create different algorithms to perform all the specific tasks. A good example is computer vision, where algorithms may need to understand if a certain object is in a picture or not.
Instead of coding the many different characteristics of that object, machine learning would start by analyzing a vast amount of pictures with that object in them, to identify similarities that would help to identify that object in the future.
Let’s make a simple example: we want to identify photos of sunny days. We would then give the algorithm a dataset with photos of clear blue skies and with photos of rainy days. The algorithm would identify the similarities of these two separate datasets, such as colors, brightness, tone, et cetera. After this learning process, the machine would be able to decide whether a photo represents a sunny day or not.
Usually, the quality of machine learning decisions is related to the size and quality of the training datasets. In fact, Machine learning algorithms heavily rely on statistics, where it is easier to make decisions upon large historical datasets.
Challenges of this technology are usually related to the lack of good training datasets, and to the capability of the algorithms to reach very high levels of accuracy required for certain applications. For example, in 2018, a self-driving car from Uber failed to detect a pedestrian, who died in the accident.
Now that you learned more about machine learning, read about its limitations here.