#style-loss

Question 1: Single choice

In neural style transfer, a layer's feature map is reshaped into $\mathbf{X}$ with one row per channel and one column per spatial position, and the style loss matches Gram matrices $G_{ij} = \sum_k x_{ik} x_{jk}$ . Why does matching Gram matrices transfer style rather than content?

Each entry sums over positions, so $G$ shows which channels co-occur, not where1
The Gram matrix keeps each channel's spatial layout, so it reproduces the brush strokes in place2
The Gram matrix is computed from the RGB pixels, so it captures only the colour palette3
The Gram matrix is normalized to unit trace, which removes the image's brightness4

Timed test