Towards encoding models for auditory cortical implants

Exploring novel approaches to auditory rehabilitation, we aim to demonstrate, in mice, the efficiency of an optogenetic cortical implant. Several studies have shown that mice can use patterned optogenetic stimulations of the sensory cortex to drive their behaviour. It was however never tested if it is possible to provide a detailed representation of sensory inputs through such stimulation patterns. To explore this key question for cortical implant devices, we developed a novel sensory encoding model based on a convolutional autoencoder, which is able to temporally compress and denoise 500ms sounds into a 10x10 array of stimulation sites while preserving latent space continuity and detailed sound information. To minimize spatial crosstalk between stimulation sites, we actually limit the latent representations to the 10 largest activations and impose spatial sparseness constraints during model training. We could then demonstrate that mice can discriminate these activity patterns when applied onto their auditory cortex using a video-projector setup for mesoscopic patterned optogenetic stimulation. After mastery of the discrimination task, we presented in catch trials various new patterns from the model and observed that several mice elicit similar behavioural categorization responses across patterns. This demonstrates that the artificial patterns imposed to auditory cortex produce a robust representation structure that can be used to solve a task. These results indicate that constrained autoencoder model can be used for generating artificial auditory perception via an array of cortical stimulators. We aim to further benchmark these artificial perceptions against already acquired auditory discrimination performances of normally-hearing mice.