Why do we use data for validation if we do not update weights for these data. Just to see how the model works on real data? So you can then see after training.
Yes, the validation accuracy gives an estimate of the model performing on unseen data.
Once your training is finished, you should test it once on a test set, as the estimate of the validation dataset might be bias, e.g. if you’ve used it for early stopping, model selection etc.