Bug Details and Reports
We detail the bug information found in this paper as follows.
Inner-computing Bug in Web Browsers
Inner-computing Bug in Web Browsers
As we stated in the paper, an obvious accuracy decline (i.e., 82.66% to 77.08%) occurs on all 3 browsers for the ResNet-20 model, after migration from PC to web platform. We carefully examined the weights/biases and outputs of each hidden layers for the discrepancy-inducing input image, and found there exists computing distinction between PC and web browsers, as shown in section IV-D in the paper.
We reported these issues to the team of TensorFlow.js (#1112)，and the developers have acknowledged that "there is a bug in the way our WebGL backend handles 1x1 conv2d kernels right now", confirming to fix in the upcoming 1.0.0 release. Below are the snapshot of our discussion.
