Any further explanation on adding noise to embedding on forward call

[po-oya]

        emb, sigma = embedded_dropout(self.encoder, torch.ones_like(self.encoder.weight), input,
                                      dropout=self.dropoute if self.training else 0,
                                      is_training=self.training)
        if self.training:
            m = torch.distributions.normal.Normal(torch.zeros_like(sigma), torch.ones_like(sigma) * 1)
            sigma = m.sample() * 0.2
            emb += sigma 

Above lines are from lm/model.py, I was going to ask what is the neccessaty of adding noise to embedding outputs? what this helps to?

Thanks