Viewing meta-learning through the optimization lens reveals strong parallels with the field of hyperparameter optimization (HPO). Both areas aim to optimize parameters that govern a learning process itself, rather than directly optimizing model parameters for a single task. This connection becomes particularly clear when we consider the bilevel optimization structure inherent in many meta-learning approaches.

The Bilevel Optimization Analogy

Recall the general form of bilevel optimization introduced earlier:

\min_{\lambda} F(\lambda, \theta^*(\lambda)) \quad \text{subject to} \quad \theta^*(\lambda) = \arg\min_{\theta} f(\lambda, \theta)

Here, $\lambda$ represents the outer-level variables, and $\theta$ represents the inner-level variables. $F$ is the outer objective, and $f$ is the inner objective.

In standard Hyperparameter Optimization (HPO), the goal is typically to find hyperparameters $\lambda$ (e.g., learning rate, regularization strength, architecture choices) that minimize a validation loss $F$ after a model $\theta^*(\lambda)$ has been trained on a training dataset using those hyperparameters. The inner loop minimizes the training loss $f$ :

Outer Variables $\lambda$ : Hyperparameters (learning rate $\alpha$ , regularization $\beta$ , etc.).
Inner Variables $\theta$ : Model parameters.
Outer Objective $F$ : Performance on a validation set (e.g., validation loss).
Inner Objective $f$ : Performance on the training set (e.g., training loss).

\min_{\lambda} \mathcal{L}_{\text{val}}(\theta^*(\lambda); \mathcal{D}_{\text{val}}) \quad \text{subject to} \quad \theta^*(\lambda) = \arg\min_{\theta} \mathcal{L}_{\text{train}}(\theta; \mathcal{D}_{\text{train}}, \lambda)

In Meta-Learning, particularly gradient-based methods like MAML, the structure is analogous. The goal is to find meta-parameters $\theta$ (often an initialization) that minimize the average loss $F$ across query sets of various tasks, after the model has been adapted $k$ steps on the corresponding support sets. The inner loop performs this task-specific adaptation, minimizing the task loss $f_i$ :

Outer Variables $\lambda$ (or Meta-Parameters $\theta_{\text{meta}}$ ): Often the initial model parameters $\theta$ , but could also include adaptation learning rates or other meta-learned components.
Inner Variables $\phi$ (or Task-Parameters $\theta'_{\text{task}}$ ): Adapted model parameters for a specific task $i$ .
Outer Objective $F$ : Average performance across tasks on query sets $\mathcal{Q}_i$ .
Inner Objective $f_i$ : Performance on the support set $\mathcal{S}_i$ for task $i$ .

For MAML, this looks like:

\min_{\theta} \mathbb{E}_{T_i \sim p(T)} [\mathcal{L}_{\mathcal{Q}_i}(\theta'_i)] \quad \text{subject to} \quad \theta'_i = \text{Adapt}(\theta, \mathcal{S}_i)

Where $\text{Adapt}(\theta, \mathcal{S}_i)$ usually involves one or more gradient descent steps on $\mathcal{L}_{\mathcal{S}_i}$ starting from $\theta$ .

The bilevel structure highlights the similarity: an outer loop optimizes parameters controlling an inner learning process, evaluated based on the outcome of that inner process.

Algorithmic Overlap

The shared bilevel structure means that algorithms developed for one domain often find application or have conceptual counterparts in the other.

Gradient-Based Methods: Techniques for computing gradients through the inner optimization process are central to both gradient-based HPO and gradient-based meta-learning. In HPO, this involves differentiating the validation loss with respect to hyperparameters, often requiring differentiation through the training process (e.g., using implicit differentiation or backpropagation through unrolled optimization steps). This mirrors the computation of the meta-gradient in MAML, which requires differentiating the query set loss with respect to the initial parameters $\theta$ through the adaptation steps performed on the support set. Algorithms like Implicit MAML (iMAML) directly leverage techniques related to implicit differentiation, also used in HPO.
Black-Box/Derivative-Free Methods: HPO often deals with hyperparameters for which gradients are unavailable or impractical to compute (e.g., discrete architectural choices). Techniques like Bayesian Optimization, Evolutionary Strategies, or Reinforcement Learning are common. While less prevalent in mainstream gradient-based meta-learning for finding initializations $\theta$ , these optimization techniques can be relevant for optimizing meta-learning hyperparameters (like the meta-learning rate, adaptation steps $k$ , network architecture choices used within the meta-learner) or in specific meta-learning contexts like meta-reinforcement learning or architecture search within meta-learning.

What Constitutes "Hyperparameters" in Meta-Learning?

From the HPO perspective, the "hyperparameters" being optimized in the meta-learning outer loop are the meta-parameters themselves. These most commonly include:

Model Initialization ( $\theta$ ): The primary target in algorithms like MAML and Reptile. The goal is to find an initialization that allows for rapid adaptation.
Adaptation Learning Rates ( $\alpha$ ): Some meta-learning approaches explicitly learn task-specific or per-parameter adaptation learning rates as part of the meta-parameters.
Learned Preprocessing/Embedding Functions: In metric-based meta-learning, the embedding network is learned in the outer loop. Its parameters act like hyperparameters governing the inner loop's distance calculations.
Meta-Optimizer Parameters: If a learned optimizer (an "optimizer function" parameterized itself) is used for adaptation, its parameters are optimized in the outer loop.

Distinctions and Considerations

Despite the parallels, there are significant differences:

Dimensionality: Meta-learning often optimizes very high-dimensional meta-parameters (e.g., the entire initialization weights $\theta$ of a foundation model), whereas HPO typically optimizes a smaller set of scalar or low-dimensional hyperparameters.
Inner Loop Structure: The inner loop in meta-learning is specifically defined by the few-shot adaptation process on a support set, often involving only a few gradient steps. The inner loop in HPO is usually a full model training process on a larger dataset.
Objective Landscape: The optimization landscapes can differ substantially. Meta-learning landscapes are influenced by the distribution of tasks and the interaction between initialization and adaptation dynamics.

Understanding the connection to HPO provides a valuable conceptual framework. It highlights that meta-learning is fundamentally about optimizing the conditions or parameters that enable effective learning, much like HPO optimizes the conditions (hyperparameters) for effective training. Techniques and insights from the mature field of HPO can inspire new approaches or provide analytical tools for understanding the behavior and challenges of optimizing meta-learning systems, especially as we scale them to large foundation models. Conversely, the techniques developed for handling high-dimensional outer loops in meta-learning might offer insights for specific HPO problems.