Interactive experiments on pretrained Conv2d layer weights.

Conv2d layers are learnable 2D signal filters, it self is indeed a function-let, so what can it actually do? 🤔

⚪ inspect into featur maps: what does the first conv2d layer do in the well-known classifiers?

⚪ inspect into conv2d kernels: what are the learned geometrical basis?

The first Conv2d layer kernels of a pretrained model:

⚪ grouping kernels: many kernels seems to be similar thus redundant?

The first Conv2d layer kernels of a pretrained model (and centroids):

kernels	kernel centroids

⚪ PGD adversarial attack: what if we attack a single conv2d layer?

⚪ mathematical property of the well-known image kernels: what are the fixed points of a 2d kernel?

• The reprocessed ImageNet-1k dataset can be downloaded here: https://pan.quark.cn/s/373b488d101e
  • NOTE: It is a subset of 1k images from validation split of original intact ImageNet dataset
• Tiny-ImageNet can be found here: tiny-imagenet-200

by Armit 2022/10/28