DeepMind, a subsidiary of Google, talked about the experiment with testing artificial intelligence models on generalization skills and abstract thinking. Specialists have developed a generator that formulates questions based on the notion of progression, color properties, shapes or sizes and their interrelationships. Similar tasks are encountered in IQ tests for people.
Most models answered questions with an accuracy of 75%. At the same time, researchers found a strict correlation between the effectiveness of tasks and the ability to identify the underlying abstractions. They managed to increase efficiency by training algorithms to explain their answers, to show what interrelations and properties should be considered in one or another issue.
However, in some models it is difficult to "transfer" the studied relationships to new properties, for example, if it trained to identify logical sequences relative to the color of objects, and in the task it is required to establish the dependence on their form.
The team found out that if the neural network correctly extrapolated its knowledge of the relationship to a new combination of values, then the accuracy of the tasks was increased to 87%. In the case of incorrect extrapolation, it fell to 32%.