Closed SonHyegang closed 7 months ago
Hi @SonHyegang, thank you for reaching out with a great question. Sorry for the late reply; it has been a hectic week for me.
Essentially, we utilize two randomly initialized neural networks as the bases, or simply x- or y- axes, for the visualizations of the loss landscape. In other words, intuitively speaking, the boundaries of these visualizations can be correspond to the random initializations of the neural networks. Likewise, the training trajectories of the neural network can be conceptualized as paths that moves from these boundaries to the center of the visualizations. I hope this helps you.
I understood that in the loss landscape visualization the z-axis is NLL. I'm curious what the x-axis and y-axis mean. Of course, we can see in
loss_landscapes.py
how the x and y values participate in the calculation, but I don't have an intuitive understanding of it.Thank you sincerely.