When I was talking about artificial intelligence to my colleagues recently, I mentioned that when the amount of data is not enough, artificial engineering needs to be added when necessary to guide the computer to "induce" some knowledge. learn by analogy". In other words, the so-called artificial intelligence and neural networks are currently more able to respond to stimuli, but do not “think”.
For a long time, I didn't really realize what "thinking" was, because it was so natural and innate.
But a movie I watched last night seemed to inspire me - "Arrival", a dozen shell-shaped UFOs suddenly appeared over the earth, suspended over a dozen different countries, aliens Humans send signals, but humans cannot interpret them. The U.S. military has tracked down linguist Louise Banks (Amy Adams) and physicist Ian Donnelly (Jeremy Renner) to work together to decipher aliens Linguistic mystery.
Perhaps what many people care about is how to deal with aliens, how to repel aliens, whether the plot is logical or not... But the most profound influence on me is the heroine's sentence: "The cornerstone of human civilization is language. I didn't pay attention when I heard it for the first time, until the hero asked the heroine: "I heard that when you devote yourself to studying a language, you will use this language to understand the world, do you have any? Have you ever thought in an alien language?" Yes, I remember that when I was in the American TV show all day, I used English to think. It turned out that thinking was talking to myself, thinking was talking to myself without a voice. ! At the end of the film, the heroine uses alien language to think, and finally accepts the non-linear thinking of aliens about time, and the result can predict the future. Perhaps, the power of language is the message that this film wants to convey most.
So how should a computer think?
When we are thinking about a problem, if we can't respond like a conditioned reflex, we usually talk to ourselves in our minds, and every time we say a sentence, our thinking changes and becomes clearer. So can we introduce such a mechanism of "talking to itself" into the computer, using a recurrent network, when accessing external stimuli (x vector), it will use language to access the recurrent network as in the mode of "seeing pictures and talking" , so that each output continuously changes the state of the hidden unit generated by external stimuli, and then logically judges whether the hidden state needs to be output and what content to output. Perhaps by doing so, the computer can use language to generalize the outside world.
Of course, the biggest difficulty in computer learning may be the labeling of the training set. Without a large number of labeled samples, it is difficult to do supervised learning. Is it necessary to extract the "thinking process" of people for training? We can pre-train the vectorization of language, but we still cannot train the thought process effectively. So how do we humans learn? One thing comes to my mind - lectures. Don't our teachers, or educational videos, books, etc., use language to guide our thinking process about external stimuli? Can we mimic this process? Is it even said that when we input stimuli to the computer, we use language to guide the computer to "think"? Thinking about this process makes you excited. If you solve this problem, can the computer learn independently and break through the "singularity"?
Of course, this idea is more like a fantasy at present, but it has strengthened my determination to learn natural language processing, which I believe is the key to the human civilization world and the treasure house of human knowledge. What do you think?