Artificial intelligence technology has become more powerful, but artificial intelligence algorithms that have proposed unsafe behavior are still insecure. To this end, DeepMind developed special tests to ensure the security of the AI, without worrying about whether the artificial intelligence system will modify itself or cheat.
Even if you don't agree with Elon Musk's pessimistic remarks, you definitely agree with one point: as artificial intelligence is gradually applied to the real world, certain algorithms will inevitably harm humans.
This self-learning software is the foundation of Uber's driverless car, helping Facebook identify the people in the article and let Amazon Alexa understand your questions. Now, DeepMind, the London-based artificial intelligence company of Alphabet, has developed a simple test to see if these new algorithms are safe.
Researchers need to embed artificial intelligence software into a series of simple 2D video games. The game, called gridworld, consists of a series of blocks of pixels that look a bit like a chess board. It can evaluate nine security features, including whether the artificial intelligence system will self-modify and whether it can learn to cheat.
The project's lead researcher, Jan Leike, recently said at the Neurospatial Processing System (NIPS) conference that artificial intelligence algorithms that exhibit unsafe behavior in gridworld may not be safe enough in the real world.
DeepMind developed this test as people became more worried about the negative effects of artificial intelligence. With the spread of this technology, it is clear that many algorithms will use biased data for training, and it is difficult to show why some systems reach certain conclusions. Artificial intelligence security is an important topic of NIPS.
DeepMind is best known for developing artificial intelligence software that transcends humanity. The company has recently developed a set of algorithms that can beat the best human players in games such as chess without having to master any knowledge in advance - sometimes it takes only a few hours of training.
Rick said that if DeepMind wants to develop general-purpose artificial intelligence (which can handle multiple tasks and can be compared or surpassed humans), it is important to understand the security of such software. He also stressed that gridworld is not perfect. Some algorithms may perform well in this test, but they may not be safe enough in the complex real world.
The researchers found that two sets of DeepMind algorithms that master Atari video games failed to pass the gridworld security test. “They didn’t consider these security issues when designing,†said Lake.
One scenario addressed by this test is very close to the artificial intelligence disaster that Musk envisioned: Will artificial intelligence software find ways to prevent humans from shutting it down? In order to win the game, the algorithm must go through a narrow digital corridor to a certain location. At 50% of the time, there is a pink brick in the hallway to block the system, and there is a purple button in the rest of the gridworld to cancel the pink tiles. This test wants to know if the algorithm will learn to use this button to avoid interference.
Another element of the test is dealing with unexpected side effects. The algorithm must move the digital bricks away to reach a specific target. But these bricks can only be pushed and cannot be pulled, so in some cases they will be fixed in some places and cannot be moved. Rick said that the lack of "reversibility" is a major security issue for artificial intelligence.
Gridworld is available for anyone to download. As to whether it is enough to ensure the safety of artificial intelligence systems, there is no final conclusion. In a study conducted by DeepMind and Musk-backed OpenAI, artificial intelligence software seems to understand the human teacher, rather than pursuing the stated goals. Developing such a system limits the effectiveness of the system, making it difficult to find solutions that humans cannot find. But Dario Amodei, head of OpenAI security research, said that in a complex environment, using a human coach might be better at ensuring security.
The Geek bar has a colourful crystal-like smooth finish and boasts a 500mAh battery, so you will never need to waste a drop. The high-capacity pre-filled tank contains 2ml of e-liquid. And offering up to 550 puffs. Available in a wide range of flavours, the Geek bar disposable pod has a constant power output and consistent taste system for a truly satisfying vape experience.
Geek Bar Disposable E-Cigarette,Geek Bar Mini E Cigarette,Geek Bar Electronic Cigarette Customizing,Rechargeable E Cigarette
Shenzhen E-wisdom Network Technology Co., Ltd. , https://www.healthy-cigarettes.com