Názov:Testing neural network intelligence using Raven's Progressive Matrices
Vedúci:prof. Ing. Igor Farkaą, Dr.
Kµúčové slová:abstract reasoning, RAVEN dataset, ResNet
Abstrakt:The thesis deals with the neural network ability to solve Raven’s Progressive Matrices, a nonverbal intelligence test widely used for estimation of human intellect. We chose ResNet-18 as the studied model, following up on the currently available results listed in other publications. Within the research, we perform both supervised and unsupervised experiments on the RAVEN dataset and support our findings with the visualization of the hidden representations on the penultimate layer of the studied model. We analyze the behavior of the network on various problem types and their combinations, also in case we eliminate certain types from the training set completely. The network demonstrates high accuracy of answers in certain experiments, which suggests that it might be capable of abstract reasoning.

Súbory bakalárskej práce:


Súbory prezentácie na obhajobe: