Visual pathways from the perspective of cost functions and multi-task deep neural networks

Area of Science:

• Computational neuroscience
• Machine learning
• Vision research

Background:

• The human visual system is organized into distinct pathways, such as the ventral (perception) and dorsal (action) streams.
• Understanding the computational principles underlying this organization is a key challenge.

Purpose of the Study:

• To computationally model the functional organization of the visual cortex using multi-task deep neural networks.
• To investigate how task relatedness influences the emergence of specialized neural representations.

Main Methods:

• Developed a novel method to measure unit contribution across multiple tasks in deep neural networks.
• Trained two deep neural networks on either related or unrelated visual tasks using identical stimuli.
• Analyzed feature representation sharing in higher-tier layers of the trained networks.

Main Results:

• Networks trained on unrelated tasks showed decreased feature sharing in higher layers.
• Networks trained on related tasks exhibited consistent high feature sharing across layers.
• Task relatedness correlated with the degree of downstream cortical-unit sharing.

Conclusions:

• The degree of task relatedness is a significant factor in the functional organization of neural systems.
• Multi-task learning in artificial neural networks can provide insights into the specialization of visual pathways.
• The proposed method can analyze neural organization in biological and artificial systems.