This neuron showed an increase of 51.5 ± 5.0 (95% confidence interval ) spikes per s per generation in response to the synthetic images and a decrease of −15.5 ± 3.5 spikes per s per generation to the reference images-thus the synthetic images became gradually more effective, despite the neuron’s slight reduction in firing rate to the reference images, presumably due to adaptation. To quantify the change in responses over time, we fit an exponential function to the cell’s mean firing rate per generation separately for the synthetic and for the reference images (solid thick lines in Figure 4A). At the beginning of the experiment, this unit responded more strongly to the reference images ( Figure S2) than to the synthetic images, but over generations, the synthetic images evolved to become more effective stimuli ( Figure 4A). The synthetic images changed with each generation as the genetic algorithm optimized the images according to the neuron’s responses ( Figure 3 first half of Video S1). We first show an example of an evolution experiment for one PIT single unit (Ri-10) in chronic-array monkey Ri. Images that were reproduced from published work with permission are as follows: “Curvature-position” (
#Xdream neural network generator#
The genetic algorithm is also able to find codes that produced images (third row) similar to the target images, indicating that not only is the generator expressive, its latent space can be searched with a genetic algorithm. To do so, we created dummy “neurons” that calculated the Euclidean distance between the target image and any given image in pixel space (left group) or CaffeNet pool5 space (right group) and used XDREAM to maximize the “neuron responses” (thereby minimizing distance to target), similar to how this network could be used to maximize firing of real neurons in electrophysiology experiments.
#Xdream neural network code#
We then asked whether, given that these images can be approximately encoded by the generator, a genetic algorithm searching in code space (“XDREAM”) is able to recover them.
The existence of codes that produced the images in the second row, regardless of how they were found, demonstrates that the deep generative network is able to encode a variety of images.
0 kommentar(er)
