AI will also “infer other things”?Baidu AI achieves milestone evolution

AsiaIndustrial NetNews: Recently, Baidu’s latest research results in the field of AI have attracted the attention of many overseas media including Quartz and Techchurch. According to reports, Xu Wei, a scientist at Baidu’s Silicon Valley Artificial Intelligence Laboratory, and his team have realized the research that allows AI to understand human natural language instructions and act in a virtual 2D environment. The research results show that the process is closer to the process of human instinctive learning. Means AI has “evolved” to the milestone of autonomous learning?

According to Quartz, the instructions used in the research results published by Baidu are word combinations that have never been used to train AI. That is to say, AI needs to fully understand the meaning of instructions and connect language with cognition and behavior. Just like parents teach their children. Baidu researchers have designed a 2D maze-like environment called XWORLD, where AI needs to learn and understand the natural language instructions of “parents” and traverse the maze under its guidance. During the training process, the researchers encouraged the correct behavior of the AI ​​in the process of crossing, and punished the wrong behavior, so as to motivate the AI ​​to understand the instructions of human natural language. More importantly, in the process, AI showed results that exceeded the researchers’ expectations, which is called “zero-shot learning” ability, which means that AI can understand unseen sentences. This research will lead us to make Machines evolve to learn like humans and go a step further.

The above picture shows the AI ​​training process of Baidu researchers. We can clearly see that the researchers guide the AI ​​path by issuing commands such as “Please move to the west side of the cabbage” and “Please move to the east side of the fig”. During the training, the researchers also asked the AI ​​simple questions about the surroundings of the environment, such as “What’s in the southeast?” to guide it in object recognition. After many training sessions, the researchers issued a new instruction “Please move to the west side of the fig”, and the AI ​​successfully completed the task. This means that even a completely new command that has never been seen before will still be able to correctly understand and execute the task. In other words, AI already has the ability to understand new instructions combined with known words and in known ways. Techchurch said that Baidu’s research is a milestone breakthrough in the field of AI.

According to Baidu, it will continue to advance the research and development of this technology from two dimensions. The first dimension is to use natural language commands in the existing 2D environment to train AI to have more capabilities and understand more humans. The second dimension is to transplant this technology from the existing 2D environment to the virtual 3D, so that AI can deal with more complex surrounding environments, and finally realize the use of natural language training by human instructors in the real environmentRobot(300024). This also gives future smart home robots,automationindustryThe development and popularization of robots, etc. bring greater room for innovation.

Following Baidu’s clear adoption of artificial intelligence as its core strategy, it has made leaps and bounds in just half a year. In addition to the evolutionary achievements of AI “inferring others” announced this time, Baidu has also integrated AI technology into its core business, subversively changing the form of human-computer interaction. For example, Baidu recently launched the DuerOS artificial intelligence operating system, which realizes the communication between humans and machines in natural language. At the same time, through the introduction of the DuerOS chip, the core functions of intelligent voice interaction are packaged with hardware chips. Open to intelligent hardware Manufacturers, promote the upgrading and development of smart home, smart travel and other fields.

The Links:   3HAC025465-011 3HAC029818-001

Published on 09/07/2022