Embodied neurocomputation in crab polarimetric vision: a deep learning approach

The semi-terrestrial crab Neohelice granulata inhabits mudflats where much of the reflected light is polarized. Previous work from our group showed that its photoreceptors process this signal through two orthogonal channels, aligned horizontally and vertically (0° and 90°, respectively). Using a deep learning approach, we trained convolutional autoencoders on a denoising task with natural scenes from the crab's habitat, comparing the biological channel (0°/90°) against a rotated control (45°/135°). The biological channel preserved the scenes' Degree of Linear Polarization (DoLP) with higher fidelity. Using Partial Information Decomposition (PID), we found that decoding the rotated channel requires 70% more synergistic processing, meaning the nervous system would need to integrate both signals jointly rather than reading them independently. Biologically, this added synergy would translate into greater processing resources and, thus, higher metabolic costs. Thus, aligning the biological channels with the 0°/90° axes would increase the efficiency of primary processing of polarimetric information.