Gradient-based meta-learning algorithms like MAML optimize parameters $\theta$ through a nested loop structure. The inner loop adapts parameters to a specific task $\mathcal{T}_i$ using its support set $D_i^{supp}$ , while the outer loop updates the meta-parameters $\theta$ based on performance across multiple tasks' query sets $D_i^{qry}$ . This bilevel optimization process, particularly when involving higher-order derivatives as in standard MAML, introduces significant challenges related to optimization stability and the variance of the meta-gradient estimate. Effectively training these models, especially large foundation models, requires careful attention to these aspects.

Optimization Instability

Instability during meta-training can manifest as exploding or vanishing gradients, leading to divergence or extremely slow progress. Several factors contribute to this:

Sharp Curvature: The loss landscapes of both the inner task-specific optimization and the outer meta-optimization can exhibit sharp curvature. This is particularly true when adapting large, complex models. High curvature can cause gradient steps to overshoot optima or oscillate wildly.
Interaction between Inner and Outer Loops: The inner loop updates directly influence the computation of the meta-gradient. If the inner loop optimization is unstable (e.g., requires many steps or a large learning rate $\alpha$ ), it can amplify noise and instability in the outer loop update.
Second-Order Derivatives (MAML): Standard MAML requires differentiating through the inner loop optimization process, involving Hessians or Hessian-vector products. Computing and utilizing these second-order terms can be computationally demanding and numerically unstable, especially in high dimensions. Small errors or approximations can accumulate, leading to poor meta-parameter updates.
Learning Rate Sensitivity: The choice of both the inner loop learning rate $\alpha$ and the outer loop learning rate $\beta$ is highly influential. An inappropriate $\alpha$ can lead to poor task adaptation, affecting the meta-gradient quality. An inappropriate $\beta$ can cause the meta-optimization itself to diverge.

High Gradient Variance

The meta-gradient is typically estimated using a mini-batch of tasks sampled from the meta-training distribution $p(\mathcal{T})$ . The gradient estimate $\nabla_\theta L_{meta}(\theta)$ often suffers from high variance due to several factors:

Task Sampling: The inherent diversity across tasks means that any finite batch of tasks provides only a noisy estimate of the true expected gradient over $p(\mathcal{T})$ .
Limited Support Data: Few-shot tasks, by definition, have small support sets $D_i^{supp}$ . The inner loop optimization based on this limited data yields task-specific parameters $\phi_i$ that are themselves somewhat noisy estimates of the optimal parameters for task $\mathcal{T}_i$ . This noise propagates to the meta-gradient calculation based on the query set $D_i^{qry}$ .
Stochasticity in Optimization: If the inner loop uses stochastic gradient descent (SGD) or involves other sources of randomness, this adds further variance to the adapted parameters $\phi_i$ and consequently to the meta-gradient.

High variance in the meta-gradient slows down convergence of the outer loop optimization, requiring smaller learning rates $\beta$ or more meta-iterations. It can also make the meta-training process less reliable, potentially converging to suboptimal meta-parameters.

Techniques for Mitigation

Addressing stability and variance is essential for successful gradient-based meta-learning. Several techniques can be employed:

1. Careful Learning Rate Management

Separate Inner and Outer Rates: Use distinct learning rates, $\alpha$ for the inner loop and $\beta$ for the outer loop. $\alpha$ controls the speed of task adaptation, while $\beta$ controls the speed of meta-learning. Tuning these independently is often necessary.
Learning Rate Scheduling: Apply learning rate decay schedules, particularly to the outer loop rate $\beta$ . A common strategy is to start with a larger $\beta$ and gradually decrease it as training progresses.
Adaptive Optimizers (Outer Loop): Using adaptive optimizers like Adam or RMSProp for the outer loop update (updating $\theta$ ) can be beneficial. Their adaptive learning rates per parameter can help navigate complex meta-loss landscapes and mitigate issues related to scaling gradients across different parameter groups. However, careful tuning of hyperparameters like $\beta_1$ , $\beta_2$ , and $\epsilon$ is still required. Using Adam for the inner loop is less common and can sometimes interfere with the meta-optimization objective, but it's not entirely ruled out.

2. Gradient Clipping

Applying gradient clipping independently to the inner loop gradients (during task adaptation) and the outer loop meta-gradients (during the meta-update) can prevent gradient explosion.

Inner Loop Clipping: Clip the gradients computed on $D_i^{supp}$ with respect to $\theta$ (or $\phi_i$ ) during the inner loop update steps.
Outer Loop Clipping: Clip the final meta-gradient computed across the batch of tasks before applying the update to $\theta$ .

Choosing appropriate clipping thresholds often requires experimentation. Norm clipping (e.g., clip_grad_norm_) is generally preferred over value clipping.

3. Meta-Optimizer Selection

While SGD with momentum or Adam are common choices for the outer loop, research explores optimizers specifically designed for bilevel problems or meta-learning. However, well-tuned standard optimizers remain prevalent and effective in many cases. The stability benefits of Adam's momentum and adaptive scaling often make it a strong baseline for the outer loop.

4. Increasing Task Batch Size

The variance of the meta-gradient estimate generally decreases as the number of tasks ( $B$ ) in the meta-batch increases. This is a direct consequence of the law of large numbers. Sampling more tasks provides a more accurate estimate of the expected gradient over the task distribution.

Illustration of how increasing the number of tasks ( $B$ ) sampled per meta-update tends to reduce the variance of the meta-gradient estimate.

The main drawback is the increased computational cost per meta-iteration, as adaptation and gradient computation must be performed for each task in the batch. Finding the right balance between variance reduction and computational feasibility is important.

5. Utilizing First-Order or Implicit Methods

FOMAML/Reptile: As discussed previously, first-order approximations like FOMAML and Reptile avoid the computation of second-order derivatives entirely. This significantly improves computational efficiency and can often enhance stability by removing the volatile Hessian terms, albeit potentially altering the optimization objective.
Implicit MAML (iMAML): iMAML computes meta-gradients using implicit differentiation, avoiding the need to unroll the inner optimization path. This can lead to constant memory requirements with respect to the number of inner steps and potentially offer better stability guarantees, especially when many inner steps are required.

6. Initialization and Regularization

Meta-Parameter Initialization: A good initialization $\theta_0$ can place the meta-parameters in a region where optimization is more stable from the start. Techniques for learning initializations are explored further in Chapter 4.
Regularization: Standard regularization techniques (e.g., L2 weight decay) applied to the meta-parameters $\theta$ can sometimes help smooth the meta-loss landscape and improve stability, although their interaction with the bilevel objective needs careful consideration.

Considerations for Foundation Models

When applying these techniques to large foundation models, the challenges are often magnified:

Higher Dimensionality: The sheer number of parameters increases the likelihood of encountering difficult optimization dynamics.
Computational Cost: Computing gradients, especially second-order ones, becomes prohibitively expensive. This makes FOMAML, Reptile, or iMAML more attractive.
Memory Constraints: Storing activations and gradients for meta-gradient computation is memory-intensive. Techniques discussed in Chapter 6 (Scaling Meta-Learning Implementations) become essential.

Therefore, combining gradient-based meta-learning with techniques like FOMAML, careful learning rate control, gradient clipping, adaptive optimizers for the outer loop, and potentially increased task batch sizes (within computational limits) forms a common strategy for stabilizing training and managing gradient variance when adapting large models.