Made with Neural Synth colab notebooks , github
techniques:
reder_naive - simple high level render, iterative gradient descent to maximize t-score to reduce distance between activations and desired activations
render_multiscale - render_naive + upsample technique to bring out features at multiple resolution
render_lapnorm - render_multiscale + laplacian normalization(cv technique) to fix high frequency noise, very little continuity, neighboring pixels have different colors, high degree of contrast
lapnorm_multi - specify multiple neurons to optimize for, and combine them with masks using numpy (matrices)
Using render-lapnorm & lapnorm_multi
experiments/trial & error to determine parameters:
h, w - output image resolution
iter_n - visibility of neurons, degree of exaggeration
picking image - compare patterns in images with patterns/feature neurons
preview + select layer & channel here
used my own images in google drive - link generator
