Deep neural networks (DNNs) are a set of machine learning algorithms whose goal is hierarchical learning and the simultaneous integration of multiple levels of abstraction for knowledge representation (Awad and Khanna, 2015). Deep learning, on which DNN systems are based, is very widespread in the field of today’s artificial intelligence systems and operates on the basis of mapping multi–layered relationships between input (independent) variables and the response variable that represents the result (López Montesinos, López Montesinos and Crossa, 2022). The question is whether DNN systems also contain the ability to solve abstract problems, which is supposed to be exclusive to human intelligence.
Deep neural networks show enormous potential in the field of visual abstraction of human IQ tests for measuring the level of abstraction. As such, the paper highlights the Dreamcoder and WReN programs, which quite effectively solve human image-cognitive tests such as Raven progressive matrices and ARC (abstract and reasoning corpus) puzzles, which cover a database of abstract and visual tasks for testing the algorithm on a wide range of tasks. DNN systems are therefore more than capable of applying abstract thinking to obtain better results, but it should be noted that they are still far from reaching the level of human abstraction.
|
