[2D5-GS-2] Machine learning: applications (1)

[2D5-GS-2-01] A Study on Developmental Artificial Neural Networks by Integrating Lower-order Functions

〇Haruka Iwai¹, Ichiro Kobayashi² (1. Information Sciences, Undergraduate School of Sciences, Ochanomizu University, 2. Advanced Sciences, Graduate School of Humanities and Sciences, Ochanomizu University)

Keywords:Developmental Artificial Neural Networks, Neural Networks, Evolutional Computation

In the conventional deep learning, the topology of the networks is determined by learning the network parameters. This leads to a lot of useless learning.
On the other hand, animals develop innate fundamental ability, called instinct, from their experiences in the environment, and acquire behavioral knowledge with higher-order functions.
Such intellectual development can be captured as "development" rather than "learning".
In this study, we adopt Weight Agnostic Neural Networks (WANN), a model that captures the topology of networks through their development, as a fundamental technology. We aim at constructing Developmental Artificial Neural Networks (DANNs) that acquire hierarchical relationships of functions, in which higher-order functions expressed from pre-existing lower-order ones.
We took up a higher order human behavior, jump forward, and tried to acquire the behavior from two pre-existing lower-order behaviors, walk and jump, applying WANN framework to developing the network including the neural networks of those two behaviors. As a result, we understood that it is necessary to more elaborate our ideas for DANN.