Editing Meta-Learning (section)

== <span style="color: #FFFFFF;">Applying</span> ==
'''MAML implementation for few-shot classification:'''
<syntaxhighlight lang="python">
import torch
import torch.nn as nn
from copy import deepcopy

class MAML:
    def __init__(self, model, inner_lr=0.01, outer_lr=0.001, n_inner_steps=5):
        self.model = model
        self.inner_lr = inner_lr
        self.optimizer = torch.optim.Adam(model.parameters(), lr=outer_lr)
        self.n_inner_steps = n_inner_steps
        self.loss_fn = nn.CrossEntropyLoss()

    def inner_adapt(self, support_x, support_y):
        """Fast adaptation on support set (simulated fine-tuning)."""
        fast_model = deepcopy(self.model)
        fast_optimizer = torch.optim.SGD(fast_model.parameters(), lr=self.inner_lr)
        for _ in range(self.n_inner_steps):
            fast_optimizer.zero_grad()
            self.loss_fn(fast_model(support_x), support_y).backward()
            fast_optimizer.step()
        return fast_model

    def meta_update(self, episodes):
        """Outer loop: update θ to minimize query loss after adaptation."""
        meta_loss = 0.0
        for support_x, support_y, query_x, query_y in episodes:
            adapted = self.inner_adapt(support_x, support_y)
            meta_loss += self.loss_fn(adapted(query_x), query_y)
        self.optimizer.zero_grad()
        (meta_loss / len(episodes)).backward()
        self.optimizer.step()
</syntaxhighlight>

; Meta-learning approach selection
: '''Image classification, few-shot''' → Prototypical Networks (simple, effective)
: '''Any architecture, gradient-based''' → MAML, FOMAML (first-order approximation)
: '''NLP, few-shot''' → In-context learning with a large LLM
: '''HPO automation''' → BOHB (Bayesian Optimization + Hyperband), Optuna
: '''NAS''' → DARTS (gradient-based), evolutionary search
</div>

<div style="background-color: #8B4500; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">