Exploiting the low-rank nature of weight updates during fine-tuning results in orders of magnitude reduction in learnable parameters