A researcher at Wageningen University recently published a study (still preprint) detailing a system by which facial recognition AI could be used to identify and measure the emotional state of farm animals .
At the moment, however, there is little evidence to believe that so-called ’emotion recognition’ systems actually work.
The system, according to the study, works like this: It is based on a data set of facial characteristics of images of farm animals collected in more than 6 farms and that has been optimized to operate with an average precision of 85% . From this, the emotional states of the animals are inferred in real time.
The software detects 13 facial actions and 9 emotional states, including whether the animal is aggressive, calm, or neutral.
According to the researcher, we are facing a very high percentage of success, but the scientific literature is still a bit skeptical of such claims.
Humans are unreliable
An AI can be trained on a data set labeled human to recognize emotions with human-like precision. However, there is no fundamental truth when it comes to human emotions .
Everyone experiences and interprets emotions differently and the way we express emotions on our faces can vary greatly based on unique and cultural biological characteristics.
In other words, it seems impossible to "train" an algorithm to recognize emotions because the training is based on human-tagged data sets. Humans make mistakes . Worse still, it is ridiculous to imagine that two humans look at a million faces and come to a highly reliable consensus on the emotional state of each person seen.
Researchers do not train AI to recognize emotions or make inferences from faces. They train artificial intelligence to mimic the perceptions of the specific humans who tagged the data they are using.
In any case, when the study is peer-reviewed and analyzed by other researchers, we’ll see if we can really tell if the pigs are happy or not :