A Study on Developmental Artificial Neural Networks that Integrate Multiple Functions using Variational Autoencoder

In conventional deep learning fine tuning or transfer learning, the synaptic weights of neural networks that have been previously learned and acquired in a particular task solution are transferred to the learning of a similar task to achieve efficient learning of that task. In this case, both tasks are assumed to be similar, and the function to be transferred to solve a task is a single one. On the other hand, organisms can abstract multiple experiences acquired in the past and adapt them as knowledge to new tasks. This requires a new neural network architecture that fuses neural networks that have learned individual problem-solving functions and is capable of transfer learning of multiple functions. To acquire such an architecture, we aim to develop a framework to fuse neural networks with multiple individual functions to acquire a new network topology. Based on this, in this study, we develop a Developmental Artificial Neural Net-works (DANN) capable of simultaneously transferring multiple functions that generate a neural network with new functions that integrate multiple functions. we use the framework of Weight Agnostic Neural Networks (WANN), in which the topology is acquired evolutionarily rather than by the synaptic weight of the networks, and the obtained topology is transformed into a matrix by using an embedded representation of the generation rules, which is then used as input to the encoder-decoder model based on Grammar Variational Autoencoder (GVAE). Through the experimental results, it was confirmed that multiple neural network models that solve individual tasks can be represented in the latent space, and that capturing the relationships among tasks.