Modelling the acquisition of symbolic concepts

December 6, 2022

Deep learning is part of our daily lives. From medical diagnosis to fraud detection, we are all subject to it without even realising it. Nevertheless, there is still a bridge to cross when talking about human-like Artificial Intelligence. These models are still not capable of learning structured knowledge, such as the manipulation of mathematical symbols without bias. There have been several attempts to build a model capable of solving arithmetic tasks. In this thesis, we investigate one of those models proposed by Cognolato and Testolin [1] capable of solving a specific arithmetic task, addition. We modified the model to test the robustness of the architecture with problems it was not built to solve, such as subtraction tasks. We perform an ablation study to analyse the impact of the halting mechanism and other parameters. The model developed for the addition was able to generalise the number of terms and the number of digits. The networks designed for the subtraction suggested that the architecture may need modifications to solve this problem. A study conducted on the halting mechanisms indicates that we need a more forgiving method to solve the subtraction problem.

Authors in the community:

• 

  B

  Beatriz de Sousa Ataíde da Silva

  ist189616

Supervisors of this institution:

• 

  M

  Mário Alexandre Teles de Figueiredo

  ist12403

electrical-engineering-electronic-engineering-information-engineering - Electrical engineering, electronic engineering, information engineering

por - Portuguese

Instituto Superior Técnico