Let $\mathcal{H}$ be the hypothetical model space defined by Few-Shot Learning, $\hat{h}$ be the model that minimize expected risk (possibly not a Few-Shot earning model), $h^{*}$ be the model in $\mathcal{H}$ that minimizes expected risk, $h_I$ be the model in $\mathcal{H}$ that minimizes empirical risk, then the total error can be decomposed as $\mathcal{E}[R(h_I)-R(\hat{h})]=\mathcal{E}[R(h^*)-R(\hat{h})]+\mathcal{E}[R(h_I)-R(h^*)]$.

The first term is Approximation Error which is affected by $\mathcal{H}$, and the second term is Estimation Error which is affected by the quantity and quality of training data.

The problem of few-shot learning is that the training sample size is small, thus causing higher estimation error, and resulting in unreliable $h_I$.