Modular Artificial Neural Networks

Jaan Erik Pihel
The work focused on the ability of modular artificial neural networks to solveproblems which are inherently divisible into smaller subtasks. The chosen problem wasto use neural networks in predicting frames of ball movements and bounces on a tablewith borders. The network had ball and border coordinates and ball speed as inputs. Aneural architecture was developed where each layer is easily interpreted and which usesattention to classify tasks. It is shown that without pre-training the learning is quite slow.However, with pre-training, the model can overfit less than different networks, becausethe model architecture has some bias and human knowledge coded into it.
Graduation Thesis language
Graduation Thesis type
Bachelor - Computer Science
Meelis Kull, Markus Kängsepp
Defence year