#time-embedding

Question 1: Single choice

In the diffusion U-Net of the Neuromatch Academy (NMA) tutorial, the time $t$ is first embedded as a vector: $\sin$ and $\cos$ of $2\pi\omega_i t$ for fixed random frequencies $\omega_i$ , followed by a learned linear layer. How does this embedding reach a convolutional feature map h of shape (B, C, H, W) inside a block?

A linear layer maps it to $C$ numbers per example, which are added at every spatial position of their channel1
It is written into the input image as an extra row of pixels before the first convolution2
It is multiplied with the convolution kernels, giving each time step its own set of weights3
It replaces the normalization statistics of each layer with values computed from $t$ 4

Timed test