Today 12:58 AM
https://www.eurekalert.org/news-releases/1135009
EXCERPT: So Adami and Gupta did an experiment where they created artificial forms of life using a computer program called Avida, and then trained an AI to detect them.
In the Avida universe, digital organisms written in computer code — essentially strings of commands — copy themselves over and over again in a virtual Petri dish inside the computer. Each time they replicate, the copying process is imperfect and their computer code changes, just like the genetic code of real organisms mutates when they reproduce.
Such forms of “digital life” have been used for several decades to study evolution, Adami explained. For the study, the researchers used Avida to generate tens of thousands of digital organisms, some of which contained the instructions needed to copy themselves and others not. They then used them to train a neural network to distinguish between the two with 99.97% accuracy.
However, when the researchers put the neural network to the test on examples it had never seen before, the results looked far less impressive.
In their experiments, the researchers started by presenting the neural network with a digital organism that the AI correctly deemed incapable of copying itself. Then by gradually swapping out one operation for another in the organism’s computer code, the researchers were able to trick the AI into misclassifying the organism as self-replicating, in as few as 150 tries.
In other words, with just a few tweaks, the team showed that it was possible to convince the AI that it was seeing signs of life where they didn’t exist.
"No matter what sequence of commands we started with, we were able to fool the AI 100% of the time," said Gupta, a PhD student in computer science and engineering at MSU. The number of sequences that could potentially trip up the network is vast. “So the likelihood of encountering such a sequence is substantial,” Adami added.
This risk of false positives raises the likelihood that space missions could be fooled into thinking they’d made a discovery by the data they gather, only to be proven wrong later.
Having trained an AI on computer-generated data, next the researchers plan to re-train the model with real-world data and see how easy it is to deceive, Gupta said. The findings underscore a known weakness in many contemporary AI models.
“AI has an Achilles heel,” said Adami, professor in MSU’s departments of microbiology and molecular genetics, and physics and astronomy. “It can see a pattern and completely misclassify it.”
Beyond the search for alien life, this weakness could also be problematic as AI continues to move into medical scanners, security cameras, self-driving cars and other devices in everyday use here on Earth... (MORE - no ads)
EXCERPT: So Adami and Gupta did an experiment where they created artificial forms of life using a computer program called Avida, and then trained an AI to detect them.
In the Avida universe, digital organisms written in computer code — essentially strings of commands — copy themselves over and over again in a virtual Petri dish inside the computer. Each time they replicate, the copying process is imperfect and their computer code changes, just like the genetic code of real organisms mutates when they reproduce.
Such forms of “digital life” have been used for several decades to study evolution, Adami explained. For the study, the researchers used Avida to generate tens of thousands of digital organisms, some of which contained the instructions needed to copy themselves and others not. They then used them to train a neural network to distinguish between the two with 99.97% accuracy.
However, when the researchers put the neural network to the test on examples it had never seen before, the results looked far less impressive.
In their experiments, the researchers started by presenting the neural network with a digital organism that the AI correctly deemed incapable of copying itself. Then by gradually swapping out one operation for another in the organism’s computer code, the researchers were able to trick the AI into misclassifying the organism as self-replicating, in as few as 150 tries.
In other words, with just a few tweaks, the team showed that it was possible to convince the AI that it was seeing signs of life where they didn’t exist.
"No matter what sequence of commands we started with, we were able to fool the AI 100% of the time," said Gupta, a PhD student in computer science and engineering at MSU. The number of sequences that could potentially trip up the network is vast. “So the likelihood of encountering such a sequence is substantial,” Adami added.
This risk of false positives raises the likelihood that space missions could be fooled into thinking they’d made a discovery by the data they gather, only to be proven wrong later.
Having trained an AI on computer-generated data, next the researchers plan to re-train the model with real-world data and see how easy it is to deceive, Gupta said. The findings underscore a known weakness in many contemporary AI models.
“AI has an Achilles heel,” said Adami, professor in MSU’s departments of microbiology and molecular genetics, and physics and astronomy. “It can see a pattern and completely misclassify it.”
Beyond the search for alien life, this weakness could also be problematic as AI continues to move into medical scanners, security cameras, self-driving cars and other devices in everyday use here on Earth... (MORE - no ads)