According to a report by the German magazine "Der Spiegel" on March 1, Italian computer scientist Simone Palazzo's artificial intelligence (AI) sometimes "dreams" about strange things. Its dreams may include an ocelot-patterned truck and a dog-headed toy car, or a reclining chair with an oversized cup of dark beer on the seat cushion. Palazzo occasionally puts his AI into a sleep mode because he hopes it will learn more thoroughly.
Sleep is crucial for human survival. When the brain enters a sleep state, it unconsciously consolidates knowledge and experiences acquired in the preceding hours and categorizes information, converting important details into long-term memory. It is hypothesized that dreams aid in storing experiences. However, researchers still cannot explain the exact function of dreams.
Unlike humans, AI's artificial neural networks never tire and do not require rest or sleep. However, they have memory deficiencies: after being taught a new task, they often forget previously learned abilities. Palazzo said, "We first train a model with images of dogs and leopards, and then show it images of pandas and giraffes, and it may suddenly fail to differentiate between dogs and leopards." Experts call this situation "catastrophic forgetting."
Palazzo and his colleague Concetto Spampinato, both from the University of Catania, are developing AI models for medical applications. They train these models to analyze X-rays of teeth or diagnose various skin diseases based on images. "Catastrophic forgetting" is a problem they face: if the AI forgets previous images when learning a new clinical image, it is of no help to anyone. Can the results of sleep research also be used to improve machines?
In their experiment, Palazzo and Spampinato simulated wake and sleep stages. First, the model was trained in a wakeful state like a regular AI: researchers input images of parrots or frogs into the neural network. Subsequently, the neural network connected artificial neurons and learned to recognize the images. This way, when it saw other images containing parrots or frogs, it could confidently identify them.
In the second step, researchers put the system into a sleep-like state through computer commands. They no longer showed it new images but only displayed images it had seen before. This repetitive process appeared as if the machine was retrieving images from memory. Palazzo said that this method can reinforce the knowledge already learned in the neural network.
Finally, in the dream mode, the neural network would see some mixed images, such as a sofa and a beer mug, or an ocelot and a truck. Palazzo hopes this will prepare the system to receive new knowledge. To achieve this, he kept the neural network in a sleep-like state for an extended period: overall, the duration of the sleep stage was about 50 times that of the wakeful stage.
People have always wanted to create machines in their own image. Palazzo also advocates for faithfully reproducing the functions of the human brain. However, this approach is controversial among computer scientists. British AI expert Andrew Rogoyski wrote in the magazine "New Scientist" that the human brain is the result of millions of years of evolution, and therefore represents a biological compromise between performance and energy consumption.
Whether Palazzo's AI is truly "sleeping" and "dreaming" may still be open to debate. As Palazzo himself admits, the neural network of his model cannot learn like a human, nor does it generate dream images on its own; rather, they are created by Italian researchers in an AI-driven image generator.
The Italian research team recently released preliminary results of the sleep experiment. Independent experts are currently reviewing the results. Depending on the model, AI networks trained in a sleep-wake cycle can achieve an increase in image recognition accuracy of up to 12%. The frequency of "catastrophic forgetting" also correspondingly decreases. Palazzo calls it "a significant improvement."
