New method for comparing neural networks exposes how artificial intelligence works



The "black box" of artificial intelligence has been examined by a team at Los Alamos National Laboratory to create a novel method for comparing neural networks that will aid researchers in understanding neural network activity. In applications like virtual assistants, facial recognition technology, and self-driving cars, neural networks are used to identify patterns in datasets.

Haydn Jones, a researcher in the Advanced Research in Cyber Systems group at Los Alamos, stated that the artificial intelligence research community "doesn't necessarily have a thorough knowledge of what neural networks are doing; they give us good outcomes, but we don't know how or why." "Our new approach performs a better job of comparing neural networks, which is key toward better understanding the mathematics behind AI," says the author.

The study "If You've Trained One You've Trained Them All: Inter-Architecture Similarity Increases With Robustness," which was recently presented at the Conference on Uncertainty in Artificial Intelligence, was written by Jones as the lead author. The paper is an important step toward characterizing the behavior of robust neural networks in addition to studying network similarity.

High performance, but fragile, neural networks. For instance, autonomous vehicles employ neural networks to recognize signs. They are fairly adept at doing this in perfect circumstances. The slightest deviation, such as a sticker on a stop sign, can, however, cause the neural network to incorrectly detect the sign and never come to a stop.

Researchers are exploring for strategies to increase network resiliency in neural networks. One cutting-edge method involves "attacking" networks as they are being trained. The AI is trained to overlook abnormalities that researchers purposefully introduce. In essence, this procedure, known as adversarial training, makes it more difficult to trick the networks.

In a surprising discovery, Jones and his collaborators from Los Alamos, Jacob Springer and Garrett Kenyon, as well as Jones' mentor Juston Moore, applied their new network similarity metric to adversarially trained neural networks. They discovered that as the severity of the attack increases, adversarial training causes neural networks in the computer vision domain to converge to very similar data representations, regardless of network architecture.

According to Jones, "we discovered that neural networks start acting the same way when we teach them to be robust against adversarial attacks."

The Los Alamos team's findings show that the addition of adversarial training significantly reduces the search area for the "correct design" for neural networks, despite the fact that there has been much effort in both industry and academia to find it. With the knowledge that adversarial training causes various designs to converge to similar answers, the AI research community may not need to invest as much time in investigating different architectures.

"Our discovery that robust neural networks resemble one another makes it simpler to comprehend how robust AI may actually function. We may even be learning more about how humans and other animals perceive things "Jones exclaimed.

DOE/Los Alamos National Laboratory

Comments

Post a Comment

Popular posts from this blog

Do You Sleep on Your Back or Side? Here's The Research on 'Optimal' Sleep Positions

Image
William Dement, a renowned sleep researcher at Stanford University, is quoted as saying that after 50 years of study, the only reason he can think of for why we sleep is "because we become sleepy." Sleep certainly matters for our health and wellbeing even though it may be, as one researcher put it, "the sole major habit in quest of a function." Do we, however, have it right? What do studies on sleeping positions say? Is there a proper sleeping position? The majority of people favor sleeping on their side. This is encouraging news because people who sleep on their backs are more likely to have trouble falling asleep or breathing during the night. Most of the time, we have a tendency to shift around quite a bit at night. According to a study involving 664 sleepers, on average, participants slept on their side 54% of the time, their back 37% of the time, and their front 7% of the time. Males are more likely to change positions throughout the night and move their arms, ...
Read more

The Science of Beards

Image
Here are a few facts that scientists have learned about beards as the world prepares to celebrate amazing facial hair on World Beard Day (September 3). Strong Manly Beard Men who have beards appear to be more dominant, aggressive, and strong, according to numerous studies[1][2][3] that have been conducted. In fact, men claim that growing facial hair makes them feel[4] more manly. Men who choose to grow beards tend to have higher testosterone levels than men who choose to keep their faces clean-shaven, according to research. They are also more likely to be hostile sexists, or to support the patriarchy and male dominance, as well as to believe that women belong in the kitchen, according to research on this topic. The beards (and muscles) on these males might be used to enforce gender roles. Professor R. D. Guthrie proposed in his book "Evolution of human threat display organs" from the 1970s that beards could be utilized to scare male rivals by improving aggressive and threaten...
Read more

Scientists Uncover a Surprising Connection Between Appetite and Sun Exposure

Image
Recent studies from Tel Aviv University show that women are not affected by sun exposure on their hunger levels, but males are. The study shows how the metabolic pathway is engaged differently in men and women. It was carried out utilizing lab models. Males of both animal and human species, as well as those exposed to the sun, are reportedly triggered to produce a protein called p53 that repairs any DNA damage that may have been done to the skin as a result of the exposure. The body produces the hunger-inducing hormone ghrelin in response to p53 activation. In females, the hormone estrogen stops p53 from interacting with ghrelin, preventing the desire to eat after being in the sun. The ground-breaking investigation was directed by professor Carmit Levy and Ph.D. candidate Shivang Parikh of the Department of Human Genetics and Biochemistry at TAU's Sackler Faculty of Medicine. In addition to the Tel Aviv Sourasky (Ichilov), Assuta, Meir, and Sheba Medical Centers, Drs. Yiftach Gepne...
Read more