Algorithm: Second-order MAML with the zeroing trick

Require: Task distribution p(T)
Require: α,η: inner loop and meta-learning rates
Require: Randomly initialized base-model parameters θ

1: Set w ← 0 (the zeroing trick)
2: while not done do
3: Sample tasks {T₁,…T_{N_batch}} from p(T)
4: for i = 1,2,…,N_batch do
5: {D_i^tr,D_i^test}← sample from T_i
6: θ_i = θ
7: for j = 1,2,…,N_step do
8: θ_i ← θ_i - α∇_{θ_i}L(θ_i,D_i^tr)
9: end for
10: end for
11: Update θ ← θ - η∇_θ ∑ _i=1^N_batchL(θ_i,D_i^test)
12: Set w ← 0 (the zeroing trick)
13: end while