Google DeepMind claims to have demonstrated that ``AI can acquire skills through social learning just like humans''
The Google DeepMind research team has announced research results that show that AI can acquire skills using the same knowledge transfer methods as humans.
Learning few-shot imitation as cultural transmission | Nature Communications
When it comes to the knowledge needed to ``act appropriately depending on the situation and achieve a goal,'' humans can recreate a new recipe they saw on TV, follow a guide on a tour, or learn how to use a copy machine from a colleague. You can efficiently acquire knowledge from other people through actions such as asking for help.
Incorporating these human knowledge transfer mechanisms into AI could eliminate the need for AI to read vast amounts of case data and use large amounts of time and computing power to acquire knowledge. Additionally, in order for humans and AI to have more useful interactions, it is beneficial for AI to know how humans communicate knowledge.
The DeepMind research team gave AI the ability to ``imitate movements'' as a means of knowledge transfer, and had it play a game of ``entering the goal in the correct order.'' By creating three situations: ``an expert accompanies you until the end,'' ``an expert accompanies you halfway,'' and ``no expert present,'' we verified whether the AI could properly learn how to play from an expert.
According to the research team, while conventional reinforcement learning-based AI could not improve scores at all, AI with new imitation ability was able to improve scores by watching and learning from the movements of experts. The left side of the figure below is an indicator of how much knowledge transfer occurred during training, and the right side is the change in score during training. In the red range of the figure, you can see that the AI is learning how to move and explore for a while after training starts, and knowledge transfer and score improvement do not occur, but when it learns to follow the expert's movements, the blue range Your score will improve rapidly as shown.
At the yellow stage, the degree of knowledge transfer is close to 1, which means that you can not only follow the movement but also be able to play the game without any problems even if the expert leaves midway. After that, as you continue training, as shown in the purple area, the degree of knowledge transfer decreases, but your score improves, indicating that you have finished absorbing the expert's knowledge and are increasing your score on your own. I understand.
The research team stated that they were able to develop a method to transfer knowledge to AI through slight imitation, and expressed their desire to develop artificial general intelligence by saying, ``The ability to transfer knowledge can play the role of an algorithm in the development of artificial general intelligence.'' Ta.
Related Posts:
