Generalized Inner Loop Meta-Learning

$$$$theta__star == argmin(theta ; L**train)

Training

$$$$def nabla(x):

x.backward()

return x.grad

theta_(t + 1) == opt_t(theta_t, varphi**opt, G_t) where G_t ==

nabla(theta_t) * ell_t**train

Training

$$$$def nabla(x):

x.backward()

return x.grad

theta__star =~ opt_T(theta_T, varphi**opt, G_T) == opt_T(P_T,

varphi**opt, nabla(theta_T) eltrain(T) (P_T, varphi**loss))

where * P_T == opt_(T - 1)(theta_(T - 1), varphi**opt, G_(T

- 1)) == opt_T - 1(P_(T - 1), varphi**opt, nabla(theta_T - 1)

eltrain(T - 1) (P_(T - 1), varphi**loss))

vdots

where * P_1 == opt_(0 *(theta_0, varphi**opt, G_0)) == opt_0(theta_0,

varphi**opt, nabla(theta_0) eltrain(0) (theta_0, varphi**loss)

)