News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
It's hard to teach robots to use language, which is why they should teach themselves. The robot is learning to process simple commands by exploring the 3D virtual world.
Devices such as amazon Alexa and Google Home have introduced voice control technology into the mainstream, but they still have to deal with simple commands. Allowing machines to intelligently handle real conversations remains a formidable challenge.
It takes a lot of new rules to try to solve this problem by directly coding the relationship between the words and objects and the action, which makes the machine unable to adapt to the new situation. This effort in machine-learning languages usually requires a great deal of human assistance.
Nowadays, DeepMind team, the Alphabet, a focus on artificial intelligence unit, together with the Carnegie Mellon university developed a kind of method, can let the machine based on the first-person shooter in the 3 d environment to find out the principle of simple language.
DevendraChaplot, a graduate student at Carnegie Mellon university, said: "doing this in 3D is definitely an important step towards success in the real world." He will present his paper at the annual meeting of the computational linguistics society. The ultimate goal, he says, is to create a simulation that is close to real life and that trained artificial intelligence can transmit what it learns to the real world.
DeepMind and Carnegie Mellon have both adopted deep reinforcement learning, popularized by DeepMind's ai technology. Neural networks get raw pixel data from virtual environments and use incentives to stimulate machines to learn through trial and error, such as high scores in games.
Usually, the goal is to get good grades in the game, but in this case, the two artificial intelligence programs has been "to the pillars of the green" such instructions, and then have to navigate to the correct object in order to obtain rewards. By running millions of training scenarios at an accelerated rate, both artificial intelligence programs have learned to associate words with specific objects and features, allowing them to execute according to instructions. They even understand the term "bigger" or "smaller" to distinguish a similar object.
Most importantly, both programs can "summarize" what they've learned and apply it to situations they've never seen before. If there are columns and red objects in the training scene, they can execute the "go to the red pillar", even though they have never seen the red column in training.
This makes them more flexible than previously rule-based systems. At Carnegie Mellon university team vision and language input are mixed together, will be the focus of artificial intelligence in the most relevant information, while DeepMind for their system provides additional learning goals, such as guess its view in the mobile will be how to change, which improve its overall performance. Because these two approaches solve this problem from different angles, their combination can provide better performance.
"These papers are preliminary, but progress is very exciting," said PedroDomingos, a professor at The university of Washington and author of The Master Algorithm. It is reported that The Master Algorithm is about different machine learning methods.
The study follows a trend in artificial intelligence, which combines difficult questions such as language and robot control. Instead, he says, it makes both challenges easier. This is because, if you have access to the real world it refers to, it is easier to understand the language, and it will be easier to understand the world through some guidance.
Millions of training means that Domingos does not believe that pure deep reinforcement learning will break the real world. He believes that AlphaGo, which is often used as a benchmark for artificial intelligence, actually shows the importance of integrating various artificial intelligence methods.
Prof MichaelLittman, a professor of intensive learning at brown university, said the results were "impressive" and that visual input was much more difficult than previous work. He noted that previous attempts to simulate ground language using simulators were limited to simple 2D environments.
But Littman responded to Domingos's concerns about the method's extensibility in the real world and pointed out that the commands were generated based on goals set by the simulator. This means that they do not really represent the inaccuracies and context-specific instructions that humans give machines in real life.
"I'm concerned that people might see examples like this," said Littman. "the network system intelligence responds to verbal commands and extrapolations, and the understanding and navigation of these online languages is much deeper than what they actually do."