Suitability of Features of Deep Convolutional Neural Networks for Modeling Somatosensory Information Processing

摘要

Deep Learning (DL) has recently led to great success in AI and attracted further attention due to the fact that the featuresextracted in its early layers have properties similar to those of real neurons in the primary visual cortex. Understandingcortical mechanisms of sensory information processing is important for improving DL systems as well as for developingmore realistic simulations of cortical systems. Using insights about how the cortex processes sensory stimuli (at least inits early areas), tactile stimulators such as Cortical Metrics have been developed as efficient tools for quantitative sensorytesting. These tools rely on computational neural models both for developing sensory tests and for mathematicallymodeling their results in clinical studies. To model the features extracted and used by the somatosensory cortex for tactileinformation processing, we propose a novel DL method using transfer learning and the principle of contextual guidance.Our approach helps describe the goal of sensory coding in early cortical areas because low-level stimulus features thatwould be behaviorally useful as building blocks enabling the construction of high-level behaviorally significant featureshave to rely on these principles that are applicable to any sensory modality (whether visual, auditory, or tactile). Applyingit to the visual domain, we easily show that our emergent features offer higher performance accuracy than AlexNet on theCaltech-101 dataset and on textures resembling tactile stimuli the primary somatosensory cortex processes. Our modelcontributes to (ⅰ) Cortical Metrics, (ⅱ) sensorimotor models, (ⅲ) deep hybrids of unsupervised and supervised networks.