深度学习表示学习特征学习与迁移学习1. 技术分析1.1 表示学习概述表示学习是自动学习数据特征的过程表示学习层次 原始数据 → 低级特征 → 中级特征 → 高级特征 → 任务预测 关键: 层次特征提取 端到端学习 迁移能力1.2 表示学习方法方法特点监督程度适用场景自监督学习无需标注无监督预训练对比学习对比正负样本自监督视觉度量学习学习距离度量弱监督检索迁移学习利用预训练模型半监督小数据1.3 迁移学习策略迁移学习策略 特征提取: 使用预训练模型特征 微调: 调整部分参数 领域适应: 适应新领域 增量学习: 逐步学习新任务2. 核心功能实现2.1 自监督学习import numpy as np class ContrastiveLearning: def __init__(self, encoder, temperature0.5): self.encoder encoder self.temperature temperature def augment(self, x): augmented x np.random.normal(0, 0.1, x.shape) return augmented def compute_loss(self, x): x1 self.augment(x) x2 self.augment(x) z1 self.encoder(x1) z2 self.encoder(x2) z np.concatenate([z1, z2], axis0) sim np.dot(z, z.T) / self.temperature sim np.exp(sim - np.max(sim, axis1, keepdimsTrue)) mask np.eye(len(z)) mask np.logical_not(mask) pos_mask np.zeros((len(z), len(z))) for i in range(len(z) // 2): pos_mask[i, i len(z) // 2] 1 pos_mask[i len(z) // 2, i] 1 pos_sim np.sum(sim * pos_mask, axis1) neg_sim np.sum(sim * mask, axis1) loss -np.log(pos_sim / neg_sim) return np.mean(loss) class SimCLR: def __init__(self, encoder, projection_dim128): self.encoder encoder self.projection_head np.random.randn(encoder.output_dim, projection_dim) def forward(self, x): features self.encoder(x) projections features self.projection_head return projections def train(self, data, epochs100): for _ in range(epochs): loss self.compute_contrastive_loss(data) self._update_parameters(loss) class MoCo: def __init__(self, encoder, queue_size65536): self.encoder_q encoder self.encoder_k encoder.copy() self.queue np.random.randn(queue_size, encoder.output_dim) self.queue_ptr 0 def enqueue(self, keys): batch_size keys.shape[0] ptr self.queue_ptr self.queue[ptr:ptrbatch_size] keys self.queue_ptr (ptr batch_size) % self.queue.shape[0]2.2 迁移学习class TransferLearning: def __init__(self, pretrained_model): self.pretrained_model pretrained_model def freeze_layers(self, num_layers_to_freeze): for i, layer in enumerate(self.pretrained_model.layers[:-num_layers_to_freeze]): for param in layer.parameters(): param.requires_grad False def unfreeze_all(self): for param in self.pretrained_model.parameters(): param.requires_grad True def replace_head(self, num_classes): in_features self.pretrained_model.classifier.in_features self.pretrained_model.classifier np.random.randn(in_features, num_classes) class FineTuning: def __init__(self, model, lr1e-5): self.model model self.lr lr def train(self, train_data, val_data, epochs10): for epoch in range(epochs): for batch in train_data: loss self._compute_loss(batch) self._update_params(loss) def _compute_loss(self, batch): X, y batch predictions self.model(X) return np.mean((predictions - y) ** 2) class DomainAdaptation: def __init__(self, source_model): self.source_model source_model def align_domains(self, source_data, target_data): source_features self.source_model.extract_features(source_data) target_features self.source_model.extract_features(target_data) self._domain_alignment_loss(source_features, target_features)2.3 度量学习class MetricLearning: def __init__(self, embedding_dim128): self.embedding_dim embedding_dim self.projection np.random.randn(embedding_dim, embedding_dim) def triplet_loss(self, anchor, positive, negative, margin1.0): anchor_emb anchor self.projection positive_emb positive self.projection negative_emb negative self.projection pos_dist np.sum((anchor_emb - positive_emb) ** 2, axis1) neg_dist np.sum((anchor_emb - negative_emb) ** 2, axis1) loss np.maximum(0, pos_dist - neg_dist margin) return np.mean(loss) class SiameseNetwork: def __init__(self, encoder): self.encoder encoder def forward(self, x1, x2): emb1 self.encoder(x1) emb2 self.encoder(x2) distance np.sum((emb1 - emb2) ** 2, axis1) return distance def contrastive_loss(self, x1, x2, y, margin1.0): distance self.forward(x1, x2) loss y * distance (1 - y) * np.maximum(0, margin - distance) return np.mean(loss)3. 性能对比3.1 自监督方法对比方法预训练效果微调效果计算成本SimCLR高高中MoCo高很高高Barlow Twins中中低3.2 迁移学习策略对比策略数据效率计算成本适用场景特征提取高低小数据集全微调中高大数据集分层微调高中中等数据集3.3 度量学习效果方法检索准确率训练难度适用场景Triplet Loss高高人脸识别Siamese中低相似度匹配ArcFace很高中大规模检索4. 最佳实践4.1 表示学习策略选择def choose_representation_learning(task_type, data_size): if data_size 1000: return transfer_learning elif task_type computer_vision: return contrastive_learning else: return self_supervised class RepresentationLearningSelector: staticmethod def select(config): strategies { contrastive: ContrastiveLearning, moco: MoCo, transfer: TransferLearning, metric: MetricLearning } return strategies[config[strategy]](**config.get(params, {}))4.2 迁移学习流程class TransferLearningWorkflow: def __init__(self): pass def run(self, pretrained_model, target_data, config): print(1. 加载预训练模型...) model self._load_pretrained(pretrained_model) print(2. 冻结底层网络...) self._freeze_layers(model, config.get(freeze_layers, 5)) print(3. 替换分类头...) self._replace_head(model, config[num_classes]) print(4. 微调...) self._fine_tune(model, target_data) return model5. 总结表示学习是深度学习的核心自监督学习无需标注学习特征对比学习通过对比学习有用特征迁移学习利用预训练模型度量学习学习距离度量对比数据如下MoCo在视觉任务上表现最好迁移学习在小数据集上最有效ArcFace是人脸识别的最佳选择推荐先尝试迁移学习再考虑自监督