Input Dimension Expandable Network: Integrating New Input Dimensions in Online Learning

You Lu Xing

Keywords: Input dimension expandable learning, bio-inspired learning, online learning, multimodal learning

Abstract: Sensory augmentation experiments have demonstrated that the perceptual dimensions of the mammalian nervous system are expandable at different levels. This capacity enables mammals to acquire signals beyond the range of their inherent sensory systems and subsequently learn to utilize such signals. A critical question arises: how to enable a learning system to expand its input dimensions in an online manner? To address this challenge, we propose a hierarchical modular neural network architecture that supports multi-level and multi-regional expansion of input dimensions, along with a dimension integration algorithm designed to guide new dimensions to proper neuron circuits during online learning. To validate our computational model, we design a series of dimension expansion experiments at different levels. The experimental results confirm that our method effectively handles the input dimension expandable learning problem.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 9109