Babies outperform artificial intelligence at detecting what motivates other people’s actions, according to new research
Babies outperform artificial intelligence at detecting what motivates other people’s actions, according to a new study by a team of psychology and data science researchers at New York University. The results, which reveal the fundamental differences between cognition and computation, point out the shortcomings of current technologies and the aspects that need to be improved so that AI can better reproduce human behavior.
Adults and even children can easily make reliable inferences about what motivates other people’s actions, something difficult for an artificial intelligence, according to the researchers. The new idea of ​​pitting babies and AI against each other in the same tasks is allowing researchers to better describe babies’ intuitive knowledge about other people and suggest ways to integrate that knowledge with artificial intelligence.
Babies are fascinated by others, as evidenced by the time they spend looking at the people around them to observe their actions and interact with them. Furthermore, previous studies focusing on the “common sense psychology” of infants (their understanding of others’ intentions, goals, preferences, and rationale behind actions) have indicated that infants are capable of assigning goals to others and expecting them to pursue them. . objectives in a rational and efficient manner. The ability to make these predictions is fundamental to human social intelligence.
On the other hand, artificial intelligence, based on machine learning algorithms, predicts actions directly. That’s why, for example, an ad promoting the Canary Islands as a tourist destination appears on your computer screen after reading a news story about Mount Teide. What AI lacks, however, is the flexibility to recognize the various contexts and situations that drive human behavior.
Babies vs. AI
To better understand the differences between human and AI skills, the researchers conducted a series of experiments with 11-month-old babies and compared their responses with those of learning-based neural network models.
To do this, they used the “Baby Intuitions Benchmark” (BIB), six common-sense psychology tasks. The BIB was designed to test the intelligence of babies and machines, allowing the performance of babies to be compared to machines and, most importantly, providing an empirical basis for building a human-like AI.
The babies watched a series of videos of simple animated shapes moving across the screen, like in a video game. The shapes’ actions simulated human behavior and decision-making by retrieving on-screen objects and other motions. Similarly, the researchers built and trained learning-based neural network models and tested the models’ responses to the exact same videos.
Their results showed that babies recognize human-like motivations even in the simplified actions of animated shapes. Infants predict that these actions are motivated by hidden but coherent goals, for example, retrieving the same object on the screen regardless of its location and moving effectively in this way even when the surrounding environment changes.
Babies demonstrate such predictions by looking longer for events that violate their predictions. The adoption of this “surprise paradigm” to study machine intelligence allows the direct comparison of an algorithm’s quantitative measure of surprise with a well-established human psychological measure: the observation time of children.
Artificial intelligence models have not been shown to understand the underlying motivations for such actions, revealing that they lack the fundamental principles of common sense psychology that babies possess.
REFERENCE