A team of scientists has demonstrated that an artificial intelligence system called a neural network can be trained to exhibit “systematic structure,” a fundamental part of human intelligence.
The research, published in the journal Nature, reveals a shift in a decades-long debate in cognitive science, a field that explores what kind of computer might best represent the human mind.
Since the 1980s, a subset of cognitive scientists has argued that neural networks, a type of artificial intelligence (AI), are not viable models because their architecture fails to capture a fundamental feature of how humans think.
But with training, neural networks can now acquire this human-like ability.
“Our work here suggests that this important aspect of human intelligence can be acquired through practice with a model that has been rejected for lacking these abilities,” says study co-author Brenden Lake, an assistant professor of psychology and data science at New York University.
Neural networks somewhat mimic the structure of the human brain, because their information processing nodes are interconnected, and their data processing flows in hierarchical layers. But historically, AI systems have not behaved like the human mind because they have lacked the ability to combine known concepts in new ways, in what is called “systematic synthesis.”
For example, Lake explained that if a standard neural network learns the words “hop,” “twice,” and “in a circle,” it should be shown many examples of how these words can be combined into meaningful phrases, such as “hop twice,” and “in a circle.” Jump in a circle." But if the system is fed a new word, such as “spin,” it again needs to see a set of examples to know how to use it similarly.
In the new study, Lake and co-author Marco Baroni of Pompeu Fabra University in Barcelona tested AI models and human volunteers using a made-up language containing words like “dax” and “wif.”
These words correspond either to colored dots, or to a function that somehow manipulates the arrangement of those dots in a particular sequence. Thus, the sequence of words determines the order in which the colored dots appear.
So, given a meaningless phrase, the AI and volunteers had to figure out basic “grammatical rules” that determine which dots go with which words.
Human participants produced the correct dot sequence about 80% of the time.
After testing seven AI models, Lake and Baroni came up with a method called meta-learning for composition (MLC), which allows a neural network to practice applying different sets of rules to newly learned words, while providing feedback on whether it applied the rules correctly. right or not.
The neural network trained by MLC matched or exceeded the performance of humans on these tests. When the researchers added data on common errors in humans, the AI model made the same types of errors as humans.
“They had impressive success at this task, in computing the meaning of sentences,” said Paul Smolensky, a professor of cognitive science at Johns Hopkins University and a senior principal investigator at Microsoft Research, who was not involved in the new study. But the model was still limited in its ability to generalize.
“He could work on the types of sentences he had been trained on, but he could not generalize them to new types of sentences,” Smolensky explained.
He added that enhancing MLC's ability to demonstrate compositional generalization is an important next step.
“Goodbye to cocaine ecstasy” with the world’s first innovative vaccine against addiction!
Scientists in Brazil announced the development of a new vaccine against addiction to cocaine and its powerful derivatives.
The experimental vaccine called "Calixcoca", which has shown promising results in animal experiments, triggers an immune response that prevents cocaine from reaching the brain, in the hope that it will help treat addiction.
If the treatment receives regulatory approval, it will be the first time cocaine addiction has been treated using a vaccine, said psychiatrist Frederico Garcia, coordinator of the team that developed the treatment at the Federal University of Minas Gerais.
The project won the top prize last week — 500,000 euros ($530,000) — at the European Health Innovation Awards for Latin American Medicine, sponsored by pharmaceutical company Eurofarma.
The vaccine works by stimulating patients' immune systems to produce antibodies that bind to cocaine molecules in the bloodstream, making them too large to pass into the brain's mesolimbic system, or "reward center," where the drug typically stimulates high levels of pleasure-inducing dopamine.
Similar studies were conducted in the United States, the world's largest consumer of cocaine, according to the United Nations Office on Drugs and Crime. But they stopped when clinical trials didn't show enough results, among other reasons, Garcia says.
To date, Calixcoca has proven effective in animal experiments, producing significant levels of antibodies against cocaine with few side effects.
Researchers also found that it protected mouse fetuses from cocaine, suggesting that it could be used in humans to protect unborn children from addicted pregnant women.
The vaccine is now scheduled to enter the final stage of trials: testing in humans.
“There is no specific registered treatment for cocaine addiction,” Garcia says. “We currently use a combination of psychological counselling, social assistance and rehabilitation, when necessary.”
He says Calixcoca could add an important tool to this system, helping patients in critical stages of recovery, such as when they leave a rehabilitation center.
He revealed that the vaccine is made of chemical compounds designed in the laboratory, rather than biological components, which means that its production will be less expensive than many vaccines and will not need to be stored at cold temperatures.
Garcia adds that it will not be a "magic cure" that can be given to anyone.