Confusion about the gradient matrix used

[MohammadHossein-Bahari]

Hello,

Thanks for the great work.
As asked before (here) I do not see why in several methods like GraNd and submodular functions, you use the concatenation of loss gradient and its multiplication with the last feature embedding as shown here:

            bias_parameters_grads = torch.autograd.grad(loss, outputs)[0] #size: batch_size,num_classes
            weight_parameters_grads = self.model.embedding_recorder.embedding.view(batch_num, 1,
                                    self.embedding_dim).repeat(1, self.args.num_classes, 1) *\
                                    bias_parameters_grads.view(batch_num, self.args.num_classes,
                                    1).repeat(1, 1, self.embedding_dim)
            gradients.append(torch.cat([bias_parameters_grads, weight_parameters_grads.flatten(1)],
                                        dim=1).cpu().numpy()) 

You are basically using the last layer features scaled by the gradient. Do you have any reasons why you choose this instead of the ones common in the literature, like gradient with respect to the last layer parameters?

Thanks!

[XianzheMa]

@PatrickZH @Chengcheng-Guo Hello we would also be curious why you didn't just use the last layer gradients but this form. Could you share with us some thoughts?