x Doesn't Seem to Change in MNIST Example

[egoetz]

I was surprised to find out that the mnist example took in images (referred to as x in the QFCModel code) with a [256, 16] format, where 256 is the batch size and 16 is the shrunken then flattened MNIST image. Why does the quantum circuit take in its examples in [1, 16] format? And how can the gates process this format if they are limited to four wires? Is every element in a given image vector fed in on the same wire or is the image truncated to the first row of its pixels?

I tried to figure out exactly how the gates were processing these images. But when I went to try and print out any changes to the input x as it progressed through encoding and the quantum layer, torch.all(torch.eq(x, old_x)) always returned true. Is x not processed until we call the measure method? If not, how do I correctly evaluate change in-between simulation steps?

[01110011011101010110010001101111]

To start answering some of your questions, I would recommend checking out the linked paper: https://arxiv.org/abs/1904.04767

From my understanding, the algorithm works by using a randomly generated trainable quantum circuit as a convolution through which groups of pixels can be slightly perturbed. Effectively, the only quantum part is the convolution (thus quanvolution). There are certainly a handful of more quantum algorithms you could explore (ex: QCNN, etc.)!