Ana Laguna Pradas had a strange desire for a first-time mother. He wanted his son to cry. His mother did not understand: "But Ana, daughter, take the child and do not let him cry anymore, he told me. It sounds like a crazy mother, but I am the first mother who wants her son to cry, "she explains.
During her pregnancy, Laguna was thinking about a question: "How am I going to understand him?" A baby communicates crying: he wants to eat, mimes, something hurts, he is sleepy. But what exactly do you want every time you cry? It was 2016, Laguna looked for apps to interpret that crying and only found a Chinese that worked badly.
"If Jane Goodall understands the language of chimpanzees, why not try to translate a newborn?"
Why not do it herself? Thought Laguna, who is a data scientist at BBVA Data & Analytics. Her intuition told her that a baby's crying has patterns and that artificial intelligence can detect them: "She had worked in machine translation, and the baby's crying is still another means of oral communication, and if Jane Goodall understands the language of chimpanzees, why not try to translate the needs of a newborn with an algorithm? "he says.
After the quarantine, he started recording his son. Each crying sample should last at least 10 seconds. That's how it was until four months. He gathered about 65 audios. In the end the little one did not cry so much: on average it turned out less than one recording a day.
The human eye sees obvious differences in audio signals, but an algorithm needs more details to find patterns. So Laguna turned to the spectograms. The intuition was confirmed: "The audio signals looked good and the accuracy of the model was acceptable," he says.
But Laguna came up with a new problem: the lack of sample, of crying. Artificial intelligence needs a substantial amount of examples.
Also, once enough examples of a type of weeping are gathered, there is another difficult question: labeling. The label is what identifies a cry like pain, hunger, sleep, desire for mimes and must be put by the parents. If the labels are misplaced, the model will look badly for the patterns and it will be a disaster. Before the algorithm can see patterns in each cry, parents must do so.
Due to lack of data and good labels, Laguna ended up dividing his database into only two options: hunger and not hunger. The initial pretension to hit with more types of tears was parked.
Now Laguna is pregnant again. This time he will be more serious. Your second child may be the Spanish baby who spends more time crying.
Laguna's only hope for increasing the database is not his second son. He has created an NGO to do data work where there is a form for other babies to collaborate with. The goal is twofold: to grow faster and avoid the bias of the entire database being the weeping of two brothers. Here Parents can raise their children's crying. Laguna wants to work with crying babies of less than 6 months.
The North American example
Laguna's intuition was also had by a Californian team led by Ariana Anderson, a computational neuropsychologist at UCLA (University of California, Los Angeles). Anderson has four children. When he heard the third cry, he began to realize that there were patterns. The next logical step was, as for Laguna, to train an algorithm to improve human perception.
Over the years, Anderson's team released an app: Chatterbaby, available on Android and iPhone, which gives a percentage with the most likely reason for crying. The Chatterbaby database has thousands of examples and is able to distinguish between pain (with 90% success), restlessness (with 85%) and hunger (with 60%). "That 60% is because we still have small samples, but since we are continuously training our algorithms with new data, precision will grow in the future," says Anderson.
The labeling certification process in Chatterbaby is delicate. The crying of pain is indisputable: they are taken out of punctures to babies by vaccines or by holes in the ear. The other two categories that at the moment handle – restless and hungry – have quality control: "They are labeled by the parents (usually the mother). Then another mother of the team and I listened one by one to those tears. If we both agree, it stays in our database. If one disagrees, it is removed. We agree on 80% crying, which shows that experienced mothers can recognize crying babies that are not theirs, "explains Anderson.
In Anderson's team, they hope to have enough data soon to identify more types of crying: separation, fear, cramps. Although each one has its difficulty. "What we understand about the normal development of babies is that they start to feel fear of separation at 6 to 7 months but often not until closer to the end of the first year," says neonatologist Diana Montoya-Williams, of Anderson's team. "Because these states do not apply to all babies in our data, we do not study it separately." For colic, we do not have enough sample yet. Preliminary investigations show that it is similar to the crying of pain with vaccines. "Chatterbaby accepts crying from children up to 2 years old, but the majority of those who have been around for 3 months.
Despite the advances of Chatterbaby, Spanish and Latin American babies must trust the work of Laguna. Children cry differently by language and country: "The baby can hear the melody (prosody) of the mother's tongue in the uterus," says Anderson. Due to the extension of the app, Chatterbaby's international database grows but it is still not enough to reach everywhere: "Around 80% of our users are international," says Anderson. By gender, however, it is "unlikely" that there are differences, adds Anderson, although they will continue to see him in his investigations.
In Google Play scores the app only has very satisfied parents who give five stars and others very disappointed, who give one. It must be very frustrating to have at 3 o'clock in the morning the latest technology in artificial intelligence and not be able to do anything while the baby continues to cry without remedy.