iMAML

iMAML, short for implicit Model-Agnostic Meta-Learning, is a gradient-based meta-learning method that aims to improve the efficiency and stability of fast adaptation to new tasks. It builds on the MAML framework but instead of differentiating through all inner-loop optimization steps, iMAML uses the implicit function theorem to compute meta-gradients. This approach reduces the memory and computational burden associated with long inner-loop unrolls.

Conceptually, iMAML treats the inner optimization as defining an optimal point w(θ) for each task, where w(θ) = argmin_w L_in(w, θ) and θ represents the meta-parameters. The outer objective evaluates performance on held-out data after adaptation, for example L_meta(θ) = E_task[L_val(w(θ); task)]. The key idea is to compute the gradient with respect to θ without backpropagating through the entire inner path. Using the implicit function theorem, the gradient involves the derivatives of the inner loss at w(θ): ∂w/∂θ = - [∂^2 L_in/∂w∂w]^{-1} [∂^2 L_in/∂w∂θ]. This allows the outer gradient to be obtained via the chain rule while circumventing explicit unrolling.

Practically, the inverse Hessian term is approximated using techniques such as conjugate gradients and Hessian-vector products, enabling scalable computation for high-dimensional models. Advantages of iMAML include reduced memory usage, the ability to employ larger inner-loop steps, and often more stable gradients compared with fully unrolled MAML. Limitations include the need to compute or approximate second-order information, which can be costly for very large models, and potential numerical challenges in solving the implicit equations.

iMAML is model-agnostic and applicable to various domains, including supervised few-shot learning and reinforcement learning, wherever MAML-style rapid adaptation is desirable.