Machine learning refers to the subfield within computer science specialized in the recognition of complex patterns in data sets. Unlike classical programming, in which a program repeatedly executes the same (more or less complex) operation, the main feature of automatic learning is that its programs manage to extract autonomously (i.e., without being specifically programmed for it) relevant information in the data being processed.
This information allows the program to “learn”, that is, to improve in its execution of the task for which it was programmed (Turing, 1950). Through the development of sophisticated algorithms (which can be understood as “models”), these approaches make it possible to identify relationships that are invisible to the human eye.
This type of algorithm interacts with us on a daily basis when, for example, the camera on our mobile phone recognizes a face or when we use an automatic translation application. In this sense, part of the success of these tools is due to their extensive field of action: from systems that detect mutations in our DNA to “big data”, which identifies patterns in huge datasets about, for example, different segments of our society.
As expected, cognitive sciences have not been immune to the development of these tools. A simple consultation of the terms “machine learning” and “brain” in a repository of articles allows us to observe that, while in 1990 only one article contained these labels, 298 works published in 2016 coincide with our search.
Following the logic of classical analyses, we would proceed to average the activity along these voxels. In all the patterns presented, there are 3 voxels activated and 3 deactivated. The average, therefore, would be the same for both animals in the two regions, suggesting that they are similarly involved in the processing of dogs and cats. However, we can observe that while in region X this inference appears to be accurate, the patterns in region Y are different, indicating a potential difference in the representation of dogs and cats in this region.
Although in this example the difference is obvious to the naked eye, these types of observations seem complicated when what we have in front of us is a broader set of data. This is where researchers have used automatic learning techniques to detect, among all the complexity of our data, patterns associated with different representations.
Basically, the functioning of these techniques consists of training an algorithm to differentiate two classes (in our case, dogs, and cats) by presenting patterns associated with both categories (yellow and blue dots in the figure). In this way, the algorithm learns which rule to use to separate specimens from each class (in this example, a diagonal line).
The key step (“test”) consists in the presentation of new unlabeled specimens, with the aim of checking the accuracy of the algorithm when assigning each specimen to the corresponding class. In this way, when the activity of a brain zone allows the specimens of two different categories to be classified above the level of chance, we will assume that this zone is representing these categories in a differential way.
What importance do these assumptions have in cognitive neuroscience? While classical techniques allowed us to detect which areas appeared to be involved in certain processes, it was not possible to understand which representations were being coded in those regions.