|
|
Rotation-equivariant convolutional neural network for design of visual prosthetic stimulation protocol
Picek, Martin ; Antolík, Ján (advisor) ; Pilát, Martin (referee)
Neighboring neurons in the primary visual cortex (V1), the first cortical area pro- cessing visual information, are selective to stimuli presented in neighboring positions of the visual field with a specific edge orientation. In this way, they form the so-called retinotopic and orientation maps of V1. Due to the absence of high-resolution cortical stimulation devices, vision restoration through prosthetic implants in V1 has not yet taken advantage of the orientation maps. However, the availability of cortical implants with stimulation resolution high enough to target separate orientation columns can be anticipated soon. Since other stimulus features are also encoded in the cortex, such as color, size, or phase, but cannot be reliably engaged even by high-resolution stimulation, in this thesis, we ask the question of how well can visual stimuli be encoded in V1 if only orientation and position preference is known. To address this question, we propose a deep neural network (DNN) providing a scalar neural activity descriptor for any targeted cortical location and multiple different orientations. This is achieved by employing a rotation-equivariant convolutional neural network (reCNN) with the last layer having only one channel for each orientation, returning the desired three-dimensional feature tensor. A...
|
guest ::
