SMP: A Cost-Effective and Efficient Method for Few-Shot Learning

We conducted a cost analysis of SMP and compared it to baseline methods. The analysis included relative adaptation time, extra inference FLOPs, and extra parameters. Results on 19 datasets from the VTAB-1k benchmark are presented in Table~ ef{table:cost}. SMP requires only 15% adaptation time on average compared to fine-tuning, while VPT requires 97%. SMP also introduces only 0.01G inference FLOPs, compared to 11.01G for VPT. While SMP requires 6.91M extra parameters on average, this is only 8% of fine-tuning. The retraining method proposed in Section~ ef{sec:retraining} can further reduce the extra parameters. Overall, SMP has a significantly lower adaptation cost, faster inference speed, and fewer extra parameters, making it an efficient method with superior performance compared to tuning-based methods.