Is neural network intelligence capable of reasoning? Or is it only based on superficial statistics? According to the qubit report, in order to detect whether artificial intelligence has reasoning ability, DeepMind recently came up with a way to do a set of IQ test questions for AI.
They built a problem generator that involved a set of abstractions, including progressive relationships, and attributes such as color and size, which produced a number of unique problems. By using the factors or combinations available to the constraint generator, researchers can create different sets of questions for training and testing the model to see how far the model can be "smart".
Through experiments, the DeepMind research team found that when the model can correctly infer the abstract concept behind the task in the test, it can produce good performance - IQ test accuracy rate can reach 87%, otherwise, the probability of answering the answer Only 32%.
The results of the study suggest that it may be unhelpful to draw general conclusions about generalization. The good or bad model that participates in the test depends on a number of factors, and in almost all cases, the system does not perform well when it needs to infer input beyond its experience, or to deal with completely unfamiliar attributes.
Foot Spa Machine,Foot Tub Spa Pedicure,Foot Bath Massager,Bubble Foot Spa
Huaian Mimir Electric Appliance Co., LTD , https://www.mimirfootbath.com