Home > News content

Google AI system DeepMind high school math exam failed

via:开源中国     time:2019/4/13 8:02:36     readed:221

Last week, Google’s DeepMind team published a book called Mathematical Reasoning for Analytic Neural Models.research report. In this study, the DeepMind team asked the AI ​​system to accept a mathematical test that included arithmetic, geometry, probability, measurement, and calculus. The questions were 40 questions, and the difficulty was about the level of high school mathematics in the UK.

Results Although DeepMind performed well on some topics, when it encountered problems such as text, addition, subtraction, multiplication and division, functions, etc., it could not answer because it could not understand the problem. Finally, in 40 questions, the artificial intelligence system only answered the 14 questions and got the score of “E”, that is, failed.

One of the topics is "1+1+1+1+1+1+1", even the pupils know that the answer is 7, but DeepMind has answered 6.

The researchers explained that the DeepMind neural model can calculate the number of occurrences (n) less than 6 times, but if n = 7, it will not come out. When the AI ​​encounters the same number multiple times, it will consider the input value to be wrong. The strange thing is that DeepMind can be figured out when it comes to longer additions. The researchers acknowledge that there is still no good explanation, but it is likely that the AI ​​neural network keeps looking at each problem and makes the right answer.

In addition, when the researchers asked DeepMind to find the "place value" in a long list of numbers, it would be fine because it could be sorted by number and rounded off.

Researchers say that human intelligence is superior to neural models in the ability to compound inferences about things. Inference is a complex, multi-faceted presentation. When answering questions, the human brain uses a variety of cognitive abilities, including classifying symbols (such as distinguishing numbers, adding, subtracting, multiplying and dividing symbols, words, variables), planning (such as finding the correct order of functions), calculating, and using working memory. To store intermediate values, you should also use the rules or theorems you learned. Conversely, DeepMind, developed with convolutional and recurrent neural networks, excels at pattern alignment, machine translation, and intensive learning, but is far less flexible than the human brain. They are not able to infer things beyond the environment of existing experience, and they are even less able to deal with information that is intentionally written.

At present, AI may not be a math teacher, but it is already quite powerful. DeepMind's AlphaGo defeated human Go champions Li Shizhen and Ke Jie in succession. Last December, DeepMind's AlphaZero won the world's top pros in a 5-0 victory in the StarCraft II test. In addition, self-study painting and creating music are also difficult to AI.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments