Toward transfer learning integrating multiple functions through the latent space

Haruka Iwai, Ichiro Kobayashi

Abstract: One notable feature of deep learning is fine-tuning. Fine-tuning is a transfer learning method in which a pre-trained model is adapted to downstream tasks by using domain-specific training data tailored to that task. However, this method focuses on transferring knowledge specialized for a single function, aiming to enhance task-solving ability. In contrast, biological systems solve complex problems by transferring multiple functions to new tasks, achieved by integrating various functions learned from different experiences. In this study, we attempt to investigate the basic principles of transfer learning of multiple functions by integrating multiple neural networks that solve a single task through a latent space. We proposed two encoding models that express neural networks in a latent space and investigated the latent representations of neural networks using each model. Through the investigation, it was confirmed that the latent representations of the proposed models well capture the characteristics of the neural network structure according to the task. In addition, in specific cases, a clear correspondence between the position in the latent space and task performance was observed.

External IDs:dblp:journals/nca/IwaiK25