University of Pennsylvania: Researchers find AI can stimulate the potential of the brain

When it comes to "black boxes", the human brain is probably the most difficult to understand. Countless scientists lament that its mechanism is too complicated to be understood by us.

Fortunately, although we cannot solve this mystery on our own, with the help of machines, the results may be different. In the latest issue of Nature Communications, a research team led by psychologist Michael Kahana of the University of Pennsylvania pointed out that the well-known "black box" - machine learning algorithms, can decode and enhance human memory. How to achieve this specifically? The answer is to send precise and real-time electrical pulses to the brain.

In other words, the researchers are going to use one "black box" to unlock the huge potential of another "black box." This approach gives people a very complicated feeling: on the one hand, it is a solution to an incredible problem; but on the other hand, it makes people feel uneasy because it sounds like the beginning of a science fiction horror movie like "apocalypse."

The best way to measure the brain is of course to go directly into the skull, but the public and relevant institutional review boards will not allow anyone to open someone else's skull under the banner of science. Therefore, the Kahana team worked with 25 epilepsy patients and implanted 100-200 electrodes in each of their brains to monitor brain discharges related to epilepsy. Through these electrodes, the Kahana team also recorded brain activity related to memory tasks.

The machine learning algorithm learned to correlate the electrode signals with the likelihood that the patient would remember the word.

Kahana's team first looked at what happens when the brain memorizes something. They collected thousands of voltage measurements per second from each electrode as the patient read and tried to remember a list of words. They then measured the patient's recall process to determine which brain activities corresponded to remembering and forgetting words.

Kahana's team repeated the above measurements several times. After measuring each patient 2-3 times, they collected enough data to train the algorithm. These algorithms can predict which words a single patient is likely to remember based on the activity of his brain electrodes.

The key to this research is that the electrodes can not only read neural activity, but also stimulate it. The researchers tried to improve (Kahana's team called it "saving") real-time memory ability by stimulating the brain. The subject will see a new word every few seconds, and the algorithm will determine whether he can remember these words based on brain signals. "This is a closed-loop system. We can record the subject's brain state, analyze whether it can remember what is in front of it, and stimulate it if it can't. The whole process only takes a few hundred milliseconds," Kahana said.

Experiments have shown that this approach is indeed effective. With the help of Kahana's team's system, these patients' memory improved by an average of 15%.

This isn’t the first time that Kahana’s lab has studied the effects of brain stimulation on memory. Last year, the team proposed that electrode pulses can both improve or worsen memory, and the key lies in the timing of the electrode pulses. In this study, the researchers found that when they stimulated the memory area of ​​the brain during the low-functioning period, the test subjects achieved higher scores (if the stimulation was applied during the high-functioning period, the effect was just the opposite). This is a major discovery, but unfortunately it is not helpful for treatment. Because researchers can only define the relationship between memory and brain state after doing memory tests. If you want to enhance memory, you must apply pulse stimulation during the memory process.

Now, Kahana and his colleagues have closed the loop with a machine learning algorithm. "It's the same algorithm that recognizes cat images, but we use it to monitor the brain's electrical activity to determine whether the brain is in a state that is conducive to learning," Kahana said. If the electrical signal shows that the brain is encoding memories efficiently, we ignore it; if not, the system quickly sends electrical pulses to the brain to make it more efficient. It acts like a brain pacemaker."

Bradley Voytek, a neuroscientist at the University of California, San Diego, commented on the research results: "It is not decisive, but it is definitely promising. The key is whether the next research in this field can produce better results. If more electrodes can be implanted in the patient's brain, the algorithm can decode more neural signals in a smaller time dimension and improve the specificity of the system. Richer data can also help improve the performance of the system. Most epilepsy patients participate in similar studies for only a few weeks. If data can be collected over a longer period of time for training, the performance of the model will definitely be better."

However, even if this system can collect more detailed and rich data, scientists must face the problem of what consequences will be produced by using "black box" algorithms to study and manipulate the human brain. Although Kahana's system can improve the subject's ability to remember words under certain circumstances, he does not know how this process is achieved. This is the characteristic of machine learning.

Fortunately, Kahana's team is aware of this, and some of the algorithms are relatively easy to examine. For this study, they used a simple linear classifier, which allowed them to draw some inferences about how the activity of individual electrodes affected the model's ability to distinguish patterns of brain activity. "We can't be sure at this point whether the features we use to record brain activity are correlated," said Youssef Ezzyat, a UPenn psychologist who led the study's machine learning analysis.

More complex deep learning techniques won’t necessarily lead to greater cognitive enhancements. But if it does work, researchers will be racking their brains to explore the principles of the algorithm and provide a bonus to the human brain with electrical pulses. Or, if it causes serious consequences, researchers will also figure out why and how to prevent it.

From: Leifeng.com