A machine learning approach for abstraction and reasoning problems without large amounts of data

Data-hungry machine learning techniques can sometimes be more successful than human intelligence for the ability they have acquired for the specific task they were trained for. However, task-based machine learning techniques with very little data are rather inconsistent compared to human cognitive abilities due to the lack of generalization. This makes it difficult for algorithms to handle volatile and hard-to-predict real life problems. Alternative approaches that have the potential to offer human-like abstraction capability are needed. This research is aimed to create an algorithm that emulates the performance-like, reasoning tasks that people apply in Intelligence Quotient (IQ) tests without the need for large amounts of data. The created algorithm solves reasoning problems in the created data set. Generalization is expected to be able to solve arbitrary complex tasks rather than a skill acquisition for a task. We obtained an accuracy score of 0.834 for the solutions created by the developed algorithm. Significance tests on the variations of accuracy have shown that consistency is achieved through unknown tasks and over-fitting problems are avoided which was not the case for task-based developed Convolutional Neural Network (CNN) methods using Cellular Automaton (CA) during this research. The algorithm on abstraction-reasoning and testing provides a benchmark for measuring Artificial Intelligence (AI) skill acquisition in unknown tasks with very small amount of data to learn.