A Biologically Inspired Neural Network for Color Representation

​

How our brain sees color is marvelous and a puzzle because color does not exist. There is light of different wavelengths that our visual system transforms into various perceptual hues. These perceived hues are contributing features in many vision-related tasks. For example, in a grocery store and looking for ripe bananas, color helps us pick the right bunch. With color as a visual feature employed in the brain, understanding how the brain transforms light to color will promote our understanding of its visual processing. Moreover, revealing color processing mechanisms in the brain can help in designing artificial vision systems. 

​

The color processing mechanisms in the visual cortex are a target of debate among color researchers. Scientists have found evidence of color processing in various visual areas in the brain and that neurons in later processing stages in the visual cortex represent a diverse set of hues. However, the existing models cannot explain the observed neuronal responses and therefore, how the brain achieves the observed color encoding is still unknown. In this work, we proposed neural network that suggests a biologically plausible transformation from light to a diverse set of hues. Computations in our network are analytically defined according to known visual processing in the brain. This network explicitly models visual areas involved in color processing and reveals the contributions of each area in enhancing the diversity of hue representation in the brain. Testing the artificial neurons in our model, we found responses of our model neurons resemble those of biological color cells. These results suggest that our model provides a novel formulation of the brain’s color processing pathway. Finally, we showed how such a biologically inspired model can be utilized in real-life applications such as flower image segmentation.

​

Paria Mehrani, Ph.D., is a Post-Doctoral Fellow at York University, where she focuses on active model learning in reinforcement learning. Prior to her post-doc, Paria was a Ph.D. candidate under the supervision of Dr. Tsotsos, where her work included developing biologically-inspired artificial systems with a focus on computer vision. During her Ph.D., she introduced analytically-defined neural networks designed according to known neural mechanisms in the brain. Her interdisciplinary work at the intersection of computer vision, machine learning, and computational neuroscience serves the dual purpose of introducing automatic vision systems and suggesting models that promote our understanding of the brain. Paria’s published work appeared in Nature Scientific Reports, ECCV, BMVC, and Vision Research. Paria was awarded a Ph.D. in Electrical Engineering and Computer Science from York University in 2021.

​

Click here for a recent Nature paper on her work.