Training on 中文AI术语词典https://terms-en.ai-term-hub.com/zh/tags/training/Recent content in Training on 中文AI术语词典Hugozh-cnSat, 18 Jul 2026 11:44:45 +0000Unslothhttps://terms-en.ai-term-hub.com/zh/terms/unsloth/Sat, 18 Jul 2026 11:37:39 +0000https://terms-en.ai-term-hub.com/zh/terms/unsloth/<h2 id="definition">Definition</h2> <p>Unsloth 是一款专为优化大语言模型（LLM）的微调和部署而设计的工具。它通过替换标准的 PyTorch 操作，实现了显著的速度提升和内存占用减少。</p> <h3 id="summary">Summary</h3> <p>Unsloth 是一个开源库，通过优化的内存管理和内核实现，将大语言模型（LLM）的训练和推理速度提升高达 2 倍。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>内存优化</li> <li>自定义内核</li> <li>LLM 微调</li> <li>加速推理</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>在有限的 GPU 资源下微调 LLM</li> <li>加速推理流水线</li> <li>降低训练时的云计算成本</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">from</span> unsloth <span style="color:#f92672">import</span> FastLanguageModel </span></span><span style="display:flex;"><span>model, tokenizer <span style="color:#f92672">=</span> FastLanguageModel<span style="color:#f92672">.</span>from_pretrained( </span></span><span style="display:flex;"><span> model_name<span style="color:#f92672">=</span><span style="color:#e6db74">&#34;unsloth/Llama-2-7b-bnb-4bit&#34;</span>, </span></span><span style="display:flex;"><span> max_seq_length<span style="color:#f92672">=</span><span style="color:#ae81ff">2048</span>, </span></span><span style="display:flex;"><span> dtype<span style="color:#f92672">=</span><span style="color:#66d9ef">None</span>, </span></span><span style="display:flex;"><span> load_in_4bit<span style="color:#f92672">=</span><span style="color:#66d9ef">True</span> </span></span><span style="display:flex;"><span>) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/lora-%E4%BD%8E%E7%A7%A9%E9%80%82%E5%BA%94/">LoRA (低秩适应)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pytorch-%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0%E6%A1%86%E6%9E%B6/">PyTorch (深度学习框架)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/hugging-face-%E6%A8%A1%E5%9E%8B%E7%A4%BE%E5%8C%BA%E5%B9%B3%E5%8F%B0/">Hugging Face (模型社区平台)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/flash-attention-%E9%AB%98%E6%95%88%E6%B3%A8%E6%84%8F%E5%8A%9B%E6%9C%BA%E5%88%B6/">Flash Attention (高效注意力机制)</a></li> </ul>人机互惠学习https://terms-en.ai-term-hub.com/zh/terms/reciprocal_human_machine_learning/Sat, 18 Jul 2026 11:31:50 +0000https://terms-en.ai-term-hub.com/zh/terms/reciprocal_human_machine_learning/<h2 id="definition">Definition</h2> <p>这种方法超越了简单的人机回环标注。它涉及双向知识转移：人类纠正模型错误，同时模型协助人类识别模式或自动化任务，从而实现持续的协同改进。</p> <h3 id="summary">Summary</h3> <p>一种协作学习范式，人类和机器通过相互教学和持续学习来提升性能。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>双向学习</li> <li>人机协作</li> <li>自适应系统</li> <li>持续改进</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>交互式标注工具</li> <li>个性化推荐调优</li> <li>专家在环的医疗诊断</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E4%BA%BA%E5%9C%A8%E5%9B%9E%E8%B7%AF-human-in-the-loop/">人在回路 (Human-in-the-loop)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E4%B8%BB%E5%8A%A8%E5%AD%A6%E4%B9%A0-active-learning/">主动学习 (Active learning)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%8D%8F%E5%90%8C%E8%BF%87%E6%BB%A4-collaborative-filtering/">协同过滤 (Collaborative filtering)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E4%BA%A4%E4%BA%92%E5%BC%8F%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0-interactive-ml/">交互式机器学习 (Interactive ML)</a></li> </ul>离线学习https://terms-en.ai-term-hub.com/zh/terms/offline_learning/Sat, 18 Jul 2026 11:28:49 +0000https://terms-en.ai-term-hub.com/zh/terms/offline_learning/<h2 id="definition">Definition</h2> <p>也称为批量学习，离线学习涉及使用先前收集的固定数据集来训练机器学习模型。与在线学习不同，模型不会根据新数据实时更新其参数，而是定期重新训练。</p> <h3 id="summary">Summary</h3> <p>离线学习是一种训练范式，模型在静态数据集上进行训练，在学习阶段不与实时环境交互。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>批量训练</li> <li>静态数据集</li> <li>模型重训练</li> <li>计算效率</li> <li>历史数据</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>基于历史用户数据训练推荐系统</li> <li>利用过往交易构建欺诈检测模型</li> <li>为存档照片开发图像分类器</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/online_learning-%E5%9C%A8%E7%BA%BF%E5%AD%A6%E4%B9%A0/">online_learning (在线学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/batch_processing-%E6%89%B9%E5%A4%84%E7%90%86/">batch_processing (批处理)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/model_training-%E6%A8%A1%E5%9E%8B%E8%AE%AD%E7%BB%83/">model_training (模型训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/data_pipeline-%E6%95%B0%E6%8D%AE%E6%B5%81%E6%B0%B4%E7%BA%BF/">data_pipeline (数据流水线)</a></li> </ul>多任务学习https://terms-en.ai-term-hub.com/zh/terms/multi_task_learning/Sat, 18 Jul 2026 11:26:49 +0000https://terms-en.ai-term-hub.com/zh/terms/multi_task_learning/<h2 id="definition">Definition</h2> <p>该技术利用相关任务之间共享的归纳偏置来提高学习效率 and 性能。通过训练单个模型同时执行多项任务，模型能够学习到更具通用性的特征表示，从而提升整体表现。</p> <h3 id="summary">Summary</h3> <p>多任务学习是一种机器学习范式，其中模型同时训练多个相关任务以提高泛化能力。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>共享表示</li> <li>归纳偏置</li> <li>任务干扰</li> <li>泛化能力</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>自然语言处理（如命名实体识别和词性标注）</li> <li>计算机视觉（如目标检测和分割）</li> <li>语音识别与合成</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer-learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">Transfer Learning (迁移学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/joint-training-%E8%81%94%E5%90%88%E8%AE%AD%E7%BB%83/">Joint Training (联合训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/neural-architecture-search-%E7%A5%9E%E7%BB%8F%E6%9E%B6%E6%9E%84%E6%90%9C%E7%B4%A2/">Neural Architecture Search (神经架构搜索)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/regularization-%E6%AD%A3%E5%88%99%E5%8C%96/">Regularization (正则化)</a></li> </ul>混合精度训练https://terms-en.ai-term-hub.com/zh/terms/mixed_precision_training/Sat, 18 Jul 2026 11:26:13 +0000https://terms-en.ai-term-hub.com/zh/terms/mixed_precision_training/<h2 id="definition">Definition</h2> <p>混合精度训练（MPT）在神经网络训练过程中结合使用半精度（FP16）和全精度（FP32）数据类型。通过使用FP16处理大多数操作，MPT减少了内存占用并提</p> <h3 id="summary">Summary</h3> <p>一种使用16位和32位浮点数进行训练的技術，旨在加速计算并减少内存使用。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>FP16</li> <li>FP32</li> <li>张量核心</li> <li>数值稳定性</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>大模型训练</li> <li>GPU加速</li> <li>内存受限环境</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">import</span> torch </span></span><span style="display:flex;"><span><span style="color:#f92672">import</span> torch.cuda.amp <span style="color:#66d9ef">as</span> amp </span></span><span style="display:flex;"><span> </span></span><span style="display:flex;"><span><span style="color:#75715e"># Example snippet showing automatic mixed precision context</span> </span></span><span style="display:flex;"><span><span style="color:#66d9ef">with</span> amp<span style="color:#f92672">.</span>autocast(): </span></span><span style="display:flex;"><span> output <span style="color:#f92672">=</span> model(input) </span></span><span style="display:flex;"><span> loss <span style="color:#f92672">=</span> criterion(output, target) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/gradient-scaling-%E6%A2%AF%E5%BA%A6%E7%BC%A9%E6%94%BE/">gradient scaling (梯度缩放)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/amp-%E8%87%AA%E5%8A%A8%E6%B7%B7%E5%90%88%E7%B2%BE%E5%BA%A6/">AMP (自动混合精度)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/half-precision-%E5%8D%8A%E7%B2%BE%E5%BA%A6/">half-precision (半精度)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/optimization-%E4%BC%98%E5%8C%96/">optimization (优化)</a></li> </ul>学习曲线https://terms-en.ai-term-hub.com/zh/terms/learning_curve/Sat, 18 Jul 2026 11:23:53 +0000https://terms-en.ai-term-hub.com/zh/terms/learning_curve/<h2 id="definition">Definition</h2> <p>通常，学习曲线在y轴上显示训练得分和验证得分，在x轴上显示训练样本数量或迭代次数。它有助于诊断模型是否存在高偏差（欠拟合）或高方差（过拟合）问题。</p> <h3 id="summary">Summary</h3> <p>学习曲线将模型的性能指标与训练数据量或训练轮数进行绘图，以可视化学习进度。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>训练得分</li> <li>验证得分</li> <li>欠拟合与过拟合</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>诊断模型性能问题</li> <li>确定样本量需求</li> <li>监控训练收敛情况</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/validation-set-%E9%AA%8C%E8%AF%81%E9%9B%86/">Validation set (验证集)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/overfitting-%E8%BF%87%E6%8B%9F%E5%90%88/">Overfitting (过拟合)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/convergence-%E6%94%B6%E6%95%9B/">Convergence (收敛)</a></li> </ul>知识蒸馏https://terms-en.ai-term-hub.com/zh/terms/knowledge_distillation/Sat, 18 Jul 2026 11:23:05 +0000https://terms-en.ai-term-hub.com/zh/terms/knowledge_distillation/<h2 id="definition">Definition</h2> <p>知识蒸馏是一种机器学习方法，用于将庞大复杂的神经网络（教师模型）压缩为更小、更高效的网络（学生模型）。学生模型经过训练以模仿教师模型的行为。</p> <h3 id="summary">Summary</h3> <p>知识蒸馏是一种模型压缩技术，其中较小的学生模型通过学习模仿较大的教师模型的行为来工作。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>师生模型</li> <li>模型压缩</li> <li>软标签</li> <li>效率</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>在边缘设备上部署模型</li> <li>降低推理延迟</li> <li>降低云计算成本</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">import</span> torch </span></span><span style="display:flex;"><span><span style="color:#f92672">import</span> torch.nn <span style="color:#66d9ef">as</span> nn </span></span><span style="display:flex;"><span> </span></span><span style="display:flex;"><span><span style="color:#66d9ef">def</span> <span style="color:#a6e22e">distillation_loss</span>(student_logits, teacher_logits, temperature<span style="color:#f92672">=</span><span style="color:#ae81ff">2.0</span>): </span></span><span style="display:flex;"><span> T <span style="color:#f92672">=</span> temperature </span></span><span style="display:flex;"><span> student_probs <span style="color:#f92672">=</span> nn<span style="color:#f92672">.</span>functional<span style="color:#f92672">.</span>softmax(student_logits <span style="color:#f92672">/</span> T, dim<span style="color:#f92672">=</span><span style="color:#ae81ff">1</span>) </span></span><span style="display:flex;"><span> teacher_probs <span style="color:#f92672">=</span> nn<span style="color:#f92672">.</span>functional<span style="color:#f92672">.</span>softmax(teacher_logits <span style="color:#f92672">/</span> T, dim<span style="color:#f92672">=</span><span style="color:#ae81ff">1</span>) </span></span><span style="display:flex;"><span> <span style="color:#66d9ef">return</span> nn<span style="color:#f92672">.</span>functional<span style="color:#f92672">.</span>kl_div( </span></span><span style="display:flex;"><span> nn<span style="color:#f92672">.</span>functional<span style="color:#f92672">.</span>log_softmax(student_logits <span style="color:#f92672">/</span> T, dim<span style="color:#f92672">=</span><span style="color:#ae81ff">1</span>), </span></span><span style="display:flex;"><span> teacher_probs, </span></span><span style="display:flex;"><span> reduction<span style="color:#f92672">=</span><span style="color:#e6db74">&#39;batchmean&#39;</span> </span></span><span style="display:flex;"><span> ) <span style="color:#f92672">*</span> (T <span style="color:#f92672">*</span> T) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/model-compression-%E6%A8%A1%E5%9E%8B%E5%8E%8B%E7%BC%A9/">Model Compression (模型压缩)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pruning-%E5%89%AA%E6%9E%9D/">Pruning (剪枝)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/quantization-%E9%87%8F%E5%8C%96/">Quantization (量化)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/neural-networks-%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C/">Neural Networks (神经网络)</a></li> </ul>Imatrixhttps://terms-en.ai-term-hub.com/zh/terms/imatrix/Sat, 18 Jul 2026 11:21:59 +0000https://terms-en.ai-term-hub.com/zh/terms/imatrix/<h2 id="definition">Definition</h2> <p>Imatrix（Importance Matrix，重要性矩阵）是一种主要与基于GGML的大语言模型（LLM）训练和量化相关的技术。它计算损失函数关于模型参数的二阶导数（即海森矩阵的近似值），从而评估每个参数对整体模型性能的重要性。这种方法有助于更精确地进行模型量化，减少精度损失，并优化微调过程中的计算效率。</p> <h3 id="summary">Summary</h3> <p>一种用于大语言模型训练的具体算法，用于计算重要性矩阵以实现高效的参数优化。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>海森矩阵</li> <li>参数重要性</li> <li>模型量化</li> <li>微调优化</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>高效的大语言模型微调</li> <li>面向边缘设备的模型量化</li> <li>降低训练过程中的计算开销</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/ggml-ggml%E5%BA%93/">GGML (GGML库)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/lora-%E4%BD%8E%E7%A7%A9%E9%80%82%E5%BA%94/">LoRA (低秩适应)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/quantization-%E9%87%8F%E5%8C%96/">Quantization (量化)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/second-order-optimization-%E4%BA%8C%E9%98%B6%E4%BC%98%E5%8C%96/">Second-Order Optimization (二阶优化)</a></li> </ul>超参数https://terms-en.ai-term-hub.com/zh/terms/hyperparameter/Sat, 18 Jul 2026 11:21:34 +0000https://terms-en.ai-term-hub.com/zh/terms/hyperparameter/<h2 id="definition">Definition</h2> <p>与模型参数（权重和偏置）不同，后者是在训练过程中从数据中学习得到的，而超参数是实践者在训练开始前选择的外部设置。它们控制着模型的结构、学习过程以及正则化强度等关键方面。</p> <h3 id="summary">Summary</h3> <p>一种在训练过程之前设定其值的配置变量，用于控制学习算法的行为。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>学习率</li> <li>批次大小</li> <li>模型架构</li> <li>正则化</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>配置神经网络训练</li> <li>设置梯度下降步骤</li> <li>定义随机森林中的树深度</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/hyperparameter_tuning-%E8%B6%85%E5%8F%82%E6%95%B0%E8%B0%83%E4%BC%98/">hyperparameter_tuning (超参数调优)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/model_parameters-%E6%A8%A1%E5%9E%8B%E5%8F%82%E6%95%B0/">model_parameters (模型参数)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/learning_rate-%E5%AD%A6%E4%B9%A0%E7%8E%87/">learning_rate (学习率)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/batch_size-%E6%89%B9%E6%AC%A1%E5%A4%A7%E5%B0%8F/">batch_size (批次大小)</a></li> </ul>超参数调优https://terms-en.ai-term-hub.com/zh/terms/hyperparameter_tuning/Sat, 18 Jul 2026 11:21:34 +0000https://terms-en.ai-term-hub.com/zh/terms/hyperparameter_tuning/<h2 id="definition">Definition</h2> <p>超参数调优涉及评估不同的超参数集，以找到能产生最高模型准确率或最低错误率的配置。常见的策略包括网格搜索、随机搜索以及基于贝叶斯优化的方法，旨在通过迭代改进来寻找最优解。</p> <h3 id="summary">Summary</h3> <p>系统地搜索最佳超参数组合以优化模型性能的过程。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>网格搜索</li> <li>随机搜索</li> <li>交叉验证</li> <li>性能指标</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>优化深度学习模型</li> <li>提高分类器准确率</li> <li>减少模型偏差和方差</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/hyperparameter_optimization-%E8%B6%85%E5%8F%82%E6%95%B0%E4%BC%98%E5%8C%96/">hyperparameter_optimization (超参数优化)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/grid_search-%E7%BD%91%E6%A0%BC%E6%90%9C%E7%B4%A2/">grid_search (网格搜索)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/random_search-%E9%9A%8F%E6%9C%BA%E6%90%9C%E7%B4%A2/">random_search (随机搜索)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/cross_validation-%E4%BA%A4%E5%8F%89%E9%AA%8C%E8%AF%81/">cross_validation (交叉验证)</a></li> </ul>顿悟学习https://terms-en.ai-term-hub.com/zh/terms/grokking/Sat, 18 Jul 2026 11:20:24 +0000https://terms-en.ai-term-hub.com/zh/terms/grokking/<h2 id="definition">Definition</h2> <p>顿悟学习（Grokking）是指深度学习中观察到的一种反直觉行为：模型在训练数据上长时间过拟合，表现出较差的泛化能力，但在经过漫长的训练周期后，突然在测试集上实现近乎完美的泛化。这种现象表明，优化过程可能存在两个截然不同的阶段：首先是记忆训练样本，随后是理解数据背后的潜在规律。顿悟学习对于理解神经网络的泛化界限、训练动力学以及记忆与学习之间的关系具有重要意义，尤其是在小数据集场景下。</p> <h3 id="summary">Summary</h3> <p>一种现象，指神经网络在小数据集上经过长时间训练后，突然展现出良好的泛化能力，远超记忆训练数据的阶段。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>延迟泛化</li> <li>过拟合</li> <li>小数据集</li> <li>优化动力学</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>研究模型泛化能力的极限</li> <li>分析训练动力学</li> <li>理解记忆与学习的区别</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/overfitting-%E8%BF%87%E6%8B%9F%E5%90%88/">overfitting (过拟合)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/generalization-%E6%B3%9B%E5%8C%96/">generalization (泛化)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/deep_learning_theory-%E6%B7%B1%E5%BA%A6%E5%AD%A6%E4%B9%A0%E7%90%86%E8%AE%BA/">deep_learning_theory (深度学习理论)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/training_dynamics-%E8%AE%AD%E7%BB%83%E5%8A%A8%E5%8A%9B%E5%AD%A6/">training_dynamics (训练动力学)</a></li> </ul>微调过的https://terms-en.ai-term-hub.com/zh/terms/finetuned/Sat, 18 Jul 2026 11:17:25 +0000https://terms-en.ai-term-hub.com/zh/terms/finetuned/<h2 id="definition">Definition</h2> <p>微调是指采用已在大型通用数据集上训练好的模型，并在较小、特定领域的数据集上继续训练的技术。这使得模型能够保留通用知识的同时，针对特定任务进行优化和适配。</p> <h3 id="summary">Summary</h3> <p>在特定数据集上进一步训练预训练模型的过程，以使其适应特定任务。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>迁移学习</li> <li>参数调整</li> <li>领域自适应</li> <li>预训练模型</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>自定义图像分类</li> <li>专用情感分析</li> <li>医疗诊断辅助</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer-learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">Transfer Learning (迁移学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pre-training-%E9%A2%84%E8%AE%AD%E7%BB%83/">Pre-training (预训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/hyperparameter-tuning-%E8%B6%85%E5%8F%82%E6%95%B0%E8%B0%83%E4%BC%98/">Hyperparameter Tuning (超参数调优)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/overfitting-%E8%BF%87%E6%8B%9F%E5%90%88/">Overfitting (过拟合)</a></li> </ul>轮次 (Epoch)https://terms-en.ai-term-hub.com/zh/terms/epoch/Sat, 18 Jul 2026 11:16:25 +0000https://terms-en.ai-term-hub.com/zh/terms/epoch/<h2 id="definition">Definition</h2> <p>在机器学习中，一个轮次代表对整个训练数据集的一次单遍迭代。在每个轮次中，模型处理所有训练样本，通过反向传播更新其权重，并评估损失函数。增加轮次数通常可以提高模型性能，但过多轮次可能导致过拟合。因此，监控验证集上的表现对于确定最佳轮次数至关重要。</p> <h3 id="summary">Summary</h3> <p>在模型训练期间，整个训练数据集通过机器学习算法的一次完整遍历。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>训练迭代</li> <li>反向传播</li> <li>收敛</li> <li>超参数调优</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>配置神经网络训练循环</li> <li>监控每个周期的验证损失</li> <li>实施早停策略</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#66d9ef">for</span> epoch <span style="color:#f92672">in</span> range(num_epochs): </span></span><span style="display:flex;"><span> <span style="color:#66d9ef">for</span> inputs, labels <span style="color:#f92672">in</span> dataloader: </span></span><span style="display:flex;"><span> optimizer<span style="color:#f92672">.</span>zero_grad() </span></span><span style="display:flex;"><span> outputs <span style="color:#f92672">=</span> model(inputs) </span></span><span style="display:flex;"><span> loss <span style="color:#f92672">=</span> criterion(outputs, labels) </span></span><span style="display:flex;"><span> loss<span style="color:#f92672">.</span>backward() </span></span><span style="display:flex;"><span> optimizer<span style="color:#f92672">.</span>step() </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/batch-size-%E6%89%B9%E6%AC%A1%E5%A4%A7%E5%B0%8F/">Batch Size (批次大小)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/iteration-%E8%BF%AD%E4%BB%A3/">Iteration (迭代)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/learning-rate-%E5%AD%A6%E4%B9%A0%E7%8E%87/">Learning Rate (学习率)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/overfitting-%E8%BF%87%E6%8B%9F%E5%90%88/">Overfitting (过拟合)</a></li> </ul>急切学习https://terms-en.ai-term-hub.com/zh/terms/eager_learning/Sat, 18 Jul 2026 11:15:46 +0000https://terms-en.ai-term-hub.com/zh/terms/eager_learning/<h2 id="definition">Definition</h2> <p>在急切学习中，系统在遇到新实例之前，会根据训练数据构建一个通用的目标函数或模型。这与延迟学习形成对比，后者将泛化过程推迟到查询阶段，即直到需要预测时才进行计算。</p> <h3 id="summary">Summary</h3> <p>急切学习是一种机器学习方法，其泛化函数在训练阶段就被学习完成，因此在训练结束后能够实现快速的预测时间。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>训练阶段泛化</li> <li>快速推理</li> <li>模型复杂度</li> <li>过拟合风险</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>实时图像分类</li> <li>欺诈检测系统</li> <li>推荐引擎</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%BB%B6%E8%BF%9F%E5%AD%A6%E4%B9%A0-%E5%B0%86%E6%B3%9B%E5%8C%96%E6%8E%A8%E8%BF%9F%E5%88%B0%E6%B5%8B%E8%AF%95%E9%98%B6%E6%AE%B5%E7%9A%84%E7%AD%96%E7%95%A5/">延迟学习 (将泛化推迟到测试阶段的策略)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%9F%BA%E4%BA%8E%E5%AE%9E%E4%BE%8B%E7%9A%84%E5%AD%A6%E4%B9%A0-%E5%A6%82knn-%E5%B1%9E%E4%BA%8E%E5%BB%B6%E8%BF%9F%E5%AD%A6%E4%B9%A0/">基于实例的学习 (如KNN，属于延迟学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0-%E4%BB%8E%E6%A0%87%E6%B3%A8%E6%95%B0%E6%8D%AE%E4%B8%AD%E5%AD%A6%E4%B9%A0%E7%9A%84%E6%96%B9%E6%B3%95/">监督学习 (从标注数据中学习的方法)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%A8%A1%E5%9E%8B%E8%AE%AD%E7%BB%83-%E8%B0%83%E6%95%B4%E6%A8%A1%E5%9E%8B%E5%8F%82%E6%95%B0%E7%9A%84%E8%BF%87%E7%A8%8B/">模型训练 (调整模型参数的过程)</a></li> </ul>早停法https://terms-en.ai-term-hub.com/zh/terms/early_stopping/Sat, 18 Jul 2026 11:15:46 +0000https://terms-en.ai-term-hub.com/zh/terms/early_stopping/<h2 id="definition">Definition</h2> <p>早停法是一种主要用于梯度下降等迭代训练过程中的正则化形式。在训练期间，模型在训练数据上的表现通常会持续改善，但在验证集上的表现可能会在某个点后开始恶化（表明出现过拟合）。早停法通过监控验证误差并在连续多个epoch未改善时停止训练来解决这一问题。</p> <h3 id="summary">Summary</h3> <p>早停法是一种正则化技术，当模型在验证集上的性能开始下降时停止训练过程，以防止过拟合。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>正则化</li> <li>验证集</li> <li>防止过拟合</li> <li>耐心参数</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>神经网络训练</li> <li>梯度提升算法</li> <li>时间序列预测模型</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/l2%E6%AD%A3%E5%88%99%E5%8C%96-%E6%9D%83%E9%87%8D%E8%A1%B0%E5%87%8F%E6%8A%80%E6%9C%AF/">L2正则化 (权重衰减技术)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/dropout-%E9%9A%8F%E6%9C%BA%E4%B8%A2%E5%BC%83%E7%A5%9E%E7%BB%8F%E5%85%83%E7%9A%84%E6%8A%80%E6%9C%AF/">Dropout (随机丢弃神经元的技术)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E4%BA%A4%E5%8F%89%E9%AA%8C%E8%AF%81-%E8%AF%84%E4%BC%B0%E6%A8%A1%E5%9E%8B%E6%B3%9B%E5%8C%96%E8%83%BD%E5%8A%9B%E7%9A%84%E6%8A%80%E6%9C%AF/">交叉验证 (评估模型泛化能力的技术)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%B3%9B%E5%8C%96%E8%AF%AF%E5%B7%AE-%E6%A8%A1%E5%9E%8B%E5%AF%B9%E6%9C%AA%E8%A7%81%E6%95%B0%E6%8D%AE%E9%A2%84%E6%B5%8B%E8%AF%AF%E5%B7%AE%E7%9A%84%E4%BC%B0%E8%AE%A1/">泛化误差 (模型对未见数据预测误差的估计)</a></li> </ul>领域自适应https://terms-en.ai-term-hub.com/zh/terms/domain_adaptation/Sat, 18 Jul 2026 11:15:33 +0000https://terms-en.ai-term-hub.com/zh/terms/domain_adaptation/<h2 id="definition">Definition</h2> <p>领域自适应解决了训练数据和测试数据来自不同分布的挑战。通过对齐有标签的源领域和无标签或稀疏标签的目标领域之间的特征表示，该方法能够提升模型的泛化能力。</p> <h3 id="summary">Summary</h3> <p>一种通过利用源领域的知识来提高模型在目标领域性能机器学习方法。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>源领域</li> <li>目标领域</li> <li>分布偏移</li> <li>特征对齐</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>跨不同扫描仪类型翻译医学图像</li> <li>适应不同口音的语音识别</li> <li>跨平台推荐系统</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer-learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">Transfer Learning (迁移学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/domain-generalization-%E9%A2%86%E5%9F%9F%E6%B3%9B%E5%8C%96/">Domain Generalization (领域泛化)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/data-augmentation-%E6%95%B0%E6%8D%AE%E5%A2%9E%E5%BC%BA/">Data Augmentation (数据增强)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/few-shot-learning-%E5%B0%91%E6%A0%B7%E6%9C%AC%E5%AD%A6%E4%B9%A0/">Few-Shot Learning (少样本学习)</a></li> </ul>宪法人工智能https://terms-en.ai-term-hub.com/zh/terms/constitutional_ai/Sat, 18 Jul 2026 11:11:15 +0000https://terms-en.ai-term-hub.com/zh/terms/constitutional_ai/<h2 id="definition">Definition</h2> <p>宪法人工智能是一种框架，旨在使大型语言模型与人类价值观保持一致，而无需在每个步骤都完全依赖人类反馈。它涉及创建一套高层级的“宪法”原则，让模型根据这些原则进行自我修正和对齐。</p> <h3 id="summary">Summary</h3> <p>一种通过基于预定义原则的自我批评，训练人工智能模型遵循安全指南的方法。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>自我纠正</li> <li>对齐</li> <li>基于原则的训练</li> <li>RLHF替代方案</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>减少大语言模型中的有害输出</li> <li>提高模型的有用性和诚实度</li> <li>高效扩展安全训练</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/rlhf-%E5%9F%BA%E4%BA%8E%E4%BA%BA%E7%B1%BB%E5%8F%8D%E9%A6%88%E7%9A%84%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0/">RLHF (基于人类反馈的强化学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/alignment-%E5%AF%B9%E9%BD%90/">alignment (对齐)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/safety-%E5%AE%89%E5%85%A8%E6%80%A7/">safety (安全性)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/llm-%E5%A4%A7%E5%9E%8B%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B/">LLM (大型语言模型)</a></li> </ul>学徒学习https://terms-en.ai-term-hub.com/zh/terms/apprenticeship_learning/Sat, 18 Jul 2026 11:05:01 +0000https://terms-en.ai-term-hub.com/zh/terms/apprenticeship_learning/<h2 id="definition">Definition</h2> <p>学徒学习，也称为基于演示的逆强化学习，使智能体能够通过观察专家行为来获取技能，而不是完全依赖奖励函数。</p> <h3 id="summary">Summary</h3> <p>一种强化学习方法，智能体通过模仿专家的演示来学习策略。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>逆强化学习</li> <li>专家演示</li> <li>策略模仿</li> <li>奖励推断</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>自动驾驶训练</li> <li>机器人操作任务</li> <li>游戏AI</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%A8%A1%E4%BB%BF%E5%AD%A6%E4%B9%A0-imitation-learning/">模仿学习 (Imitation Learning)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0-reinforcement-learning/">强化学习 (Reinforcement Learning)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%9F%BA%E4%BA%8E%E6%BC%94%E7%A4%BA%E7%9A%84%E5%AD%A6%E4%B9%A0-demonstration-based-learning/">基于演示的学习 (Demonstration-Based Learning)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%A5%96%E5%8A%B1%E5%A1%91%E9%80%A0-reward-shaping/">奖励塑造 (Reward Shaping)</a></li> </ul>对抗机器学习https://terms-en.ai-term-hub.com/zh/terms/adversarial_machine_learning/Sat, 18 Jul 2026 11:04:26 +0000https://terms-en.ai-term-hub.com/zh/terms/adversarial_machine_learning/<h2 id="definition">Definition</h2> <p>该领域既包括破坏模型的进攻性技术，也包括加固模型的防御策略。它涉及在对抗样本上训练模型以提高其韧性，这一过程被称为对抗训练。</p> <h3 id="summary">Summary</h3> <p>对抗机器学习是一个研究领域，旨在研究如何使机器学习模型对旨在欺骗它们的恶意输入保持鲁棒性。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>对抗训练</li> <li>鲁棒性</li> <li>泛化差距</li> <li>安全性</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>提高模型韧性</li> <li>AI系统的安全审计</li> <li>开发防御机制</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/adversarial_attack-%E5%AF%B9%E6%8A%97%E6%94%BB%E5%87%BB/">adversarial_attack (对抗攻击)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/defense_mechanisms-%E9%98%B2%E5%BE%A1%E6%9C%BA%E5%88%B6/">defense_mechanisms (防御机制)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/robustness-%E9%B2%81%E6%A3%92%E6%80%A7/">robustness (鲁棒性)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/privacy_preserving_ml-%E9%9A%90%E7%A7%81%E4%BF%9D%E6%8A%A4%E6%9C%BA%E5%99%A8%E5%AD%A6%E4%B9%A0/">privacy_preserving_ml (隐私保护机器学习)</a></li> </ul>监督微调https://terms-en.ai-term-hub.com/zh/terms/supervised_fine_tuning/Sat, 18 Jul 2026 11:02:04 +0000https://terms-en.ai-term-hub.com/zh/terms/supervised_fine_tuning/<h2 id="definition">Definition</h2> <p>监督微调（SFT）涉及采用大型预训练模型（如语言模型），并在较小的高质量、针对特定下游任务标注的数据集上继续训练该模型。</p> <h3 id="summary">Summary</h3> <p>在特定数据集上进一步训练预训练模型，使其适应特定任务或领域的过程。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>预训练模型</li> <li>迁移学习</li> <li>指令微调</li> <li>领域适配</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>定制聊天机器人开发</li> <li>专业医疗问答系统</li> <li>代码生成助手</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span>model<span style="color:#f92672">.</span>train() </span></span><span style="display:flex;"><span><span style="color:#66d9ef">for</span> batch <span style="color:#f92672">in</span> dataloader: </span></span><span style="display:flex;"><span> inputs, labels <span style="color:#f92672">=</span> batch </span></span><span style="display:flex;"><span> outputs <span style="color:#f92672">=</span> model(inputs, labels<span style="color:#f92672">=</span>labels) </span></span><span style="display:flex;"><span> loss <span style="color:#f92672">=</span> outputs<span style="color:#f92672">.</span>loss </span></span><span style="display:flex;"><span> loss<span style="color:#f92672">.</span>backward() </span></span><span style="display:flex;"><span> optimizer<span style="color:#f92672">.</span>step() </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E9%A2%84%E8%AE%AD%E7%BB%83-pre-training/">预训练 (Pre-training)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%9F%BA%E4%BA%8E%E4%BA%BA%E7%B1%BB%E5%8F%8D%E9%A6%88%E7%9A%84%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0-rlhf/">基于人类反馈的强化学习 (RLHF)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%8F%90%E7%A4%BA%E5%B7%A5%E7%A8%8B-prompt-engineering/">提示工程 (Prompt Engineering)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%A4%A7%E8%AF%AD%E8%A8%80%E6%A8%A1%E5%9E%8B-llm/">大语言模型 (LLM)</a></li> </ul>监督学习https://terms-en.ai-term-hub.com/zh/terms/supervised_learning/Sat, 18 Jul 2026 11:02:04 +0000https://terms-en.ai-term-hub.com/zh/terms/supervised_learning/<h2 id="definition">Definition</h2> <p>在监督学习中，算法在带标签的数据集上进行训练，意味着每个输入示例都与正确的输出配对。目标是让模型学习输入与输出之间的潜在关系。</p> <h3 id="summary">Summary</h3> <p>一种机器学习范式，模型基于带标签的训练示例学习将输入映射到输出。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>带标签数据</li> <li>输入-输出映射</li> <li>分类</li> <li>回归</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>垃圾邮件检测</li> <li>房价预测</li> <li>图像识别</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">from</span> sklearn.linear_model <span style="color:#f92672">import</span> LinearRegression </span></span><span style="display:flex;"><span>model <span style="color:#f92672">=</span> LinearRegression() </span></span><span style="display:flex;"><span>model<span style="color:#f92672">.</span>fit(X_train, y_train) </span></span><span style="display:flex;"><span>predictions <span style="color:#f92672">=</span> model<span style="color:#f92672">.</span>predict(X_test) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%97%A0%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0/">无监督学习</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E8%AE%AD%E7%BB%83%E9%9B%86/">训练集</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E9%AA%8C%E8%AF%81%E9%9B%86/">验证集</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%A8%A1%E5%9E%8B%E5%87%86%E7%A1%AE%E7%8E%87/">模型准确率</a></li> </ul>自监督学习https://terms-en.ai-term-hub.com/zh/terms/self_supervised_learning/Sat, 18 Jul 2026 11:01:52 +0000https://terms-en.ai-term-hub.com/zh/terms/self_supervised_learning/<h2 id="definition">Definition</h2> <p>自监督学习是一种技术，算法从无标签数据本身创建监督信号，通常通过预测输入的缺失部分来实现。它弥合了无监督学习和&hellip;</p> <h3 id="summary">Summary</h3> <p>一种训练方法，模型从输入数据中生成自己的标签以学习表示。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>预训练</li> <li>掩码语言建模</li> <li>对比学习</li> <li>表示学习</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>训练大型语言模型</li> <li>图像表示学习</li> <li>语音识别系统</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span>null </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pre_training-%E9%A2%84%E8%AE%AD%E7%BB%83/">pre_training (预训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/unsupervised_learning-%E6%97%A0%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0/">unsupervised_learning (无监督学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transformer-transformer/">transformer (Transformer)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/contrastive_loss-%E5%AF%B9%E6%AF%94%E6%8D%9F%E5%A4%B1/">contrastive_loss (对比损失)</a></li> </ul>损失函数https://terms-en.ai-term-hub.com/zh/terms/loss_function/Sat, 18 Jul 2026 11:00:35 +0000https://terms-en.ai-term-hub.com/zh/terms/loss_function/<h2 id="definition">Definition</h2> <p>损失函数也被称为成本函数或误差函数，它提供一个标量值，指示模型的执行表现。在训练过程中，优化算法利用该值来计算梯度，从而更新模型参数以最小化误差。</p> <h3 id="summary">Summary</h3> <p>在训练期间量化预测值与实际目标值之间差异的数学函数。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>反向传播</li> <li>梯度计算</li> <li>优化</li> <li>误差指标</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>训练监督学习模型</li> <li>评估模型性能</li> <li>超参数调优</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">import</span> torch.nn <span style="color:#66d9ef">as</span> nn </span></span><span style="display:flex;"><span>criterion <span style="color:#f92672">=</span> nn<span style="color:#f92672">.</span>CrossEntropyLoss() </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/backpropagation-%E5%8F%8D%E5%90%91%E4%BC%A0%E6%92%AD/">backpropagation (反向传播)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/gradient_descent-%E6%A2%AF%E5%BA%A6%E4%B8%8B%E9%99%8D/">gradient_descent (梯度下降)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/cross_entropy-%E4%BA%A4%E5%8F%89%E7%86%B5/">cross_entropy (交叉熵)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/mse-%E5%9D%87%E6%96%B9%E8%AF%AF%E5%B7%AE/">mse (均方误差)</a></li> </ul>学习率https://terms-en.ai-term-hub.com/zh/terms/learning_rate/Sat, 18 Jul 2026 11:00:35 +0000https://terms-en.ai-term-hub.com/zh/terms/learning_rate/<h2 id="definition">Definition</h2> <p>学习率决定了在每次训练迭代中，模型权重相对于计算出的梯度更新了多少。如果学习率过高，可能导致模型在优化过程中越过最优解；如果过低，则可能导致训练收敛缓慢。</p> <h3 id="summary">Summary</h3> <p>控制模型优化过程中步长的超参数，旨在最小化损失函数。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>梯度下降</li> <li>超参数调优</li> <li>收敛</li> <li>优化器</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>神经网络训练</li> <li>深度学习模型优化</li> <li>强化学习策略更新</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">import</span> torch.optim <span style="color:#66d9ef">as</span> optim </span></span><span style="display:flex;"><span>model <span style="color:#f92672">=</span> MyModel() </span></span><span style="display:flex;"><span>optimizer <span style="color:#f92672">=</span> optim<span style="color:#f92672">.</span>SGD(model<span style="color:#f92672">.</span>parameters(), lr<span style="color:#f92672">=</span><span style="color:#ae81ff">0.01</span>) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/gradient_descent-%E6%A2%AF%E5%BA%A6%E4%B8%8B%E9%99%8D/">gradient_descent (梯度下降)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/optimizer-%E4%BC%98%E5%8C%96%E5%99%A8/">optimizer (优化器)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/hyperparameter-%E8%B6%85%E5%8F%82%E6%95%B0/">hyperparameter (超参数)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/convergence-%E6%94%B6%E6%95%9B/">convergence (收敛)</a></li> </ul>梯度下降https://terms-en.ai-term-hub.com/zh/terms/gradient_descent/Sat, 18 Jul 2026 11:00:15 +0000https://terms-en.ai-term-hub.com/zh/terms/gradient_descent/<h2 id="definition">Definition</h2> <p>梯度下降是一种用于寻找可微函数局部最小值的一阶迭代优化算法。在机器学习中，它沿梯度的反方向更新模型权重，以逐步降低损失。</p> <h3 id="summary">Summary</h3> <p>一种迭代优化算法，通过调整模型参数来最小化损失函数。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>损失函数</li> <li>学习率</li> <li>优化</li> <li>反向传播</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>训练深度神经网络</li> <li>线性回归参数调优</li> <li>图像识别模型优化</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/backpropagation-%E5%8F%8D%E5%90%91%E4%BC%A0%E6%92%AD/">Backpropagation (反向传播)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/learning-rate-%E5%AD%A6%E4%B9%A0%E7%8E%87/">Learning Rate (学习率)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/optimizer-%E4%BC%98%E5%8C%96%E5%99%A8/">Optimizer (优化器)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/loss-function-%E6%8D%9F%E5%A4%B1%E5%87%BD%E6%95%B0/">Loss Function (损失函数)</a></li> </ul>少样本学习https://terms-en.ai-term-hub.com/zh/terms/few_shot_learning/Sat, 18 Jul 2026 11:00:03 +0000https://terms-en.ai-term-hub.com/zh/terms/few_shot_learning/<h2 id="definition">Definition</h2> <p>少样本学习旨在使模型能够从寥寥几个示例中进行泛化，模仿人类的学习效率。它通常依赖于元学习策略，即模型先在一系列相关任务上进行预训练，从而获得快速适应新任务的能力，即使这些新任务只有少量标注数据可用。</p> <h3 id="summary">Summary</h3> <p>少样本学习是一种机器学习范式，模型仅从非常有限的标记训练数据中学习新概念。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>元学习</li> <li>数据效率</li> <li>泛化能力</li> <li>任务适配</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>罕见疾病识别</li> <li>新产品分类</li> <li>自定义物体检测</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/meta_learning-%E5%85%83%E5%AD%A6%E4%B9%A0/">meta_learning (元学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer_learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">transfer_learning (迁移学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/data_efficiency-%E6%95%B0%E6%8D%AE%E6%95%88%E7%8E%87/">data_efficiency (数据效率)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/supervised_learning-%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0/">supervised_learning (监督学习)</a></li> </ul>微调后的https://terms-en.ai-term-hub.com/zh/terms/fine_tuned/Sat, 18 Jul 2026 10:56:23 +0000https://terms-en.ai-term-hub.com/zh/terms/fine_tuned/<h2 id="definition">Definition</h2> <p>微调涉及使用较小且针对特定任务的数据集，继续训练已在大型通用数据集上训练好的模型。该技术利用了预训练模型中习得的通用特征表示，使其能够以较低的计算成本和较少的数据量，高效地适配到具体的应用场景中。</p> <h3 id="summary">Summary</h3> <p>在特定数据集上进一步训练预训练模型，使其适应特定下游任务的过程。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>迁移学习</li> <li>权重更新</li> <li>任务特定</li> <li>预训练模型</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>调整大语言模型以审阅法律文档</li> <li>定制视觉模型用于工业缺陷检测</li> <li>专门化语音识别以适应特定口音</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pre_training-%E9%A2%84%E8%AE%AD%E7%BB%83/">pre_training (预训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer_learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">transfer_learning (迁移学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/supervised_fine_tuning-%E7%9B%91%E7%9D%A3%E5%BE%AE%E8%B0%83/">supervised_fine_tuning (监督微调)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/parameter_efficient-%E5%8F%82%E6%95%B0%E9%AB%98%E6%95%88/">parameter_efficient (参数高效)</a></li> </ul>监督式https://terms-en.ai-term-hub.com/zh/terms/supervised/Sat, 18 Jul 2026 10:55:02 +0000https://terms-en.ai-term-hub.com/zh/terms/supervised/<h2 id="definition">Definition</h2> <p>监督学习涉及向算法提供包含输入和正确答案（标签）的数据。模型通过最小化预测误差来学习将输入映射到输出。这项技术&hellip;</p> <h3 id="summary">Summary</h3> <p>一种机器学习范式，模型在有标签的输入-输出对上进行训练。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>有标签数据</li> <li>映射</li> <li>损失最小化</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>图像分类</li> <li>垃圾邮件检测</li> <li>价格预测</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">from</span> sklearn.linear_model <span style="color:#f92672">import</span> LinearRegression </span></span><span style="display:flex;"><span>model <span style="color:#f92672">=</span> LinearRegression() </span></span><span style="display:flex;"><span>model<span style="color:#f92672">.</span>fit(X_train, y_train) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%97%A0%E7%9B%91%E7%9D%A3-unsupervised/">无监督 (Unsupervised)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%A0%87%E7%AD%BE-label/">标签 (Label)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E5%9B%9E%E5%BD%92-regression/">回归 (Regression)</a></li> </ul>预训练https://terms-en.ai-term-hub.com/zh/terms/pre_training/Sat, 18 Jul 2026 10:53:51 +0000https://terms-en.ai-term-hub.com/zh/terms/pre_training/<h2 id="definition">Definition</h2> <p>预训练是深度学习中的一种基础技术，模型从海量数据中学习广泛的特征和模式，通常无需标签。这一过程使模型能够发展出通用的知识表示，从而在后续针对特定下游任务进行微调时，仅需少量数据即可达到高性能。</p> <h3 id="summary">Summary</h3> <p>在大型未标记数据集上训练机器学习模型的初始阶段，以便在针对特定任务进行微调之前学习通用表示。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>迁移学习</li> <li>特征提取</li> <li>大规模数据</li> <li>微调</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>训练BERT或GPT等语言模型</li> <li>使用ImageNet权重初始化卷积神经网络（CNN）</li> <li>构建多模态AI的基础模型</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">from</span> transformers <span style="color:#f92672">import</span> BertModel </span></span><span style="display:flex;"><span>model <span style="color:#f92672">=</span> BertModel<span style="color:#f92672">.</span>from_pretrained(<span style="color:#e6db74">&#39;bert-base-uncased&#39;</span>) </span></span><span style="display:flex;"><span><span style="color:#75715e"># Model is now pre-trained and ready for fine-tuning</span> </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/fine-tuning-%E5%BE%AE%E8%B0%83/">Fine-tuning (微调)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/foundation-model-%E5%9F%BA%E7%A1%80%E6%A8%A1%E5%9E%8B/">Foundation Model (基础模型)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/unsupervised-learning-%E6%97%A0%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0/">Unsupervised Learning (无监督学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer-learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">Transfer Learning (迁移学习)</a></li> </ul>损失https://terms-en.ai-term-hub.com/zh/terms/loss/Sat, 18 Jul 2026 10:52:50 +0000https://terms-en.ai-term-hub.com/zh/terms/loss/<h2 id="definition">Definition</h2> <p>损失函数（也称为成本函数）衡量机器学习模型的预测结果与真实标签在训练期间的匹配程度。优化算法的目标是最小化这个损失值，从而提高模型的准确性。</p> <h3 id="summary">Summary</h3> <p>一个数值，用于量化模型预测值与实际目标值之间的误差。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>成本函数</li> <li>优化</li> <li>梯度下降</li> <li>误差指标</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>训练图像分类器</li> <li>优化回归模型</li> <li>评估模型收敛性</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">import</span> torch.nn <span style="color:#66d9ef">as</span> nn </span></span><span style="display:flex;"><span>criterion <span style="color:#f92672">=</span> nn<span style="color:#f92672">.</span>CrossEntropyLoss() </span></span><span style="display:flex;"><span>loss <span style="color:#f92672">=</span> criterion(outputs, targets) </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/accuracy-%E5%87%86%E7%A1%AE%E7%8E%87/">Accuracy (准确率)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/gradient-descent-%E6%A2%AF%E5%BA%A6%E4%B8%8B%E9%99%8D/">Gradient Descent (梯度下降)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/cross-entropy-%E4%BA%A4%E5%8F%89%E7%86%B5/">Cross-Entropy (交叉熵)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/overfitting-%E8%BF%87%E6%8B%9F%E5%90%88/">Overfitting (过拟合)</a></li> </ul>指令微调https://terms-en.ai-term-hub.com/zh/terms/instruction_tuning/Sat, 18 Jul 2026 10:52:05 +0000https://terms-en.ai-term-hub.com/zh/terms/instruction_tuning/<h2 id="definition">Definition</h2> <p>这一过程弥合了通用预训练与特定任务表现之间的差距。通过让模型接触多样化的指令-响应对，它学会了泛化到未见过的任务，而无需针对每个新任务进行额外的调整（注：原文截断，此处补全语义）。这是使大语言模型具备人类对齐能力的关键步骤。</p> <h3 id="summary">Summary</h3> <p>指令微调是一种微调技术，通过在包含指令及其对应响应的数据集上训练预训练语言模型，以提升其遵循任务指令的能力。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>微调</li> <li>监督学习</li> <li>零样本泛化</li> <li>人类对齐</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>构建聊天机器人</li> <li>提高代码生成的准确性</li> <li>使模型符合安全准则</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/fine-tuning-%E5%BE%AE%E8%B0%83/">fine-tuning (微调)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/rlhf-%E5%9F%BA%E4%BA%8E%E4%BA%BA%E7%B1%BB%E5%8F%8D%E9%A6%88%E7%9A%84%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0/">RLHF (基于人类反馈的强化学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pre-training-%E9%A2%84%E8%AE%AD%E7%BB%83/">pre-training (预训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/alignment-%E5%AF%B9%E9%BD%90/">alignment (对齐)</a></li> </ul>微调https://terms-en.ai-term-hub.com/zh/terms/fine/Sat, 18 Jul 2026 10:51:19 +0000https://terms-en.ai-term-hub.com/zh/terms/fine/<h2 id="definition">Definition</h2> <p>微调涉及采用在大型数据集上训练过的通用模型，并在较小、更专业的数据集上进行进一步训练，以提高其在特定任务上的性能。该技术利用了迁移学习的优势。</p> <h3 id="summary">Summary</h3> <p>微调是指使用额外数据将预训练的AI模型适应特定任务或领域的过程。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>迁移学习</li> <li>预训练模型</li> <li>领域适应</li> <li>权重调整</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>医学影像诊断</li> <li>法律文档分析</li> <li>自定义情感分析</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/transfer-learning-%E8%BF%81%E7%A7%BB%E5%AD%A6%E4%B9%A0/">Transfer Learning (迁移学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/pre-training-%E9%A2%84%E8%AE%AD%E7%BB%83/">Pre-training (预训练)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/adaptation-%E9%80%82%E5%BA%94/">Adaptation (适应)</a></li> </ul>微调https://terms-en.ai-term-hub.com/zh/terms/fine_tuning/Sat, 18 Jul 2026 07:44:46 +0000https://terms-en.ai-term-hub.com/zh/terms/fine_tuning/<h2 id="definition">Definition</h2> <p>微调涉及在一个已在大而通用数据集上训练好的模型基础上，继续在专业化数据集上进行训练。这使得模型在保留通用知识的同时，能够习得特定任务的技能。</p> <h3 id="summary">Summary</h3> <p>使用较小的数据集将预训练模型适配到特定下游任务的过程。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>迁移学习</li> <li>预训练模型</li> <li>任务特定适配</li> <li>学习率</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>适配大语言模型以构建客户服务聊天机器人</li> <li>专门化图像分类器用于医疗诊断</li> <li>定制语音识别以针对特定口音</li> </ul> <h2 id="code-example">Code Example</h2> <div class="highlight"><pre tabindex="0" style="color:#f8f8f2;background-color:#272822;-moz-tab-size:4;-o-tab-size:4;tab-size:4;"><code class="language-python" data-lang="python"><span style="display:flex;"><span><span style="color:#f92672">from</span> transformers <span style="color:#f92672">import</span> AutoModelForSequenceClassification </span></span><span style="display:flex;"><span>model <span style="color:#f92672">=</span> AutoModelForSequenceClassification<span style="color:#f92672">.</span>from_pretrained(<span style="color:#e6db74">&#39;bert-base-uncased&#39;</span>) </span></span><span style="display:flex;"><span><span style="color:#75715e"># Freeze base layers</span> </span></span><span style="display:flex;"><span><span style="color:#66d9ef">for</span> param <span style="color:#f92672">in</span> model<span style="color:#f92672">.</span>bert<span style="color:#f92672">.</span>parameters(): </span></span><span style="display:flex;"><span> param<span style="color:#f92672">.</span>requires_grad <span style="color:#f92672">=</span> <span style="color:#66d9ef">False</span> </span></span><span style="display:flex;"><span><span style="color:#75715e"># Train only classification head</span> </span></span></code></pre></div><h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E9%A2%84%E8%AE%AD%E7%BB%83/">预训练</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E6%8F%90%E7%A4%BA%E5%B7%A5%E7%A8%8B/">提示工程</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/lora-%E4%BD%8E%E7%A7%A9%E9%80%82%E5%BA%94/">LoRA (低秩适应)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/%E7%9B%91%E7%9D%A3%E5%AD%A6%E4%B9%A0/">监督学习</a></li> </ul>反向传播https://terms-en.ai-term-hub.com/zh/terms/backpropagation/Sat, 18 Jul 2026 07:44:10 +0000https://terms-en.ai-term-hub.com/zh/terms/backpropagation/<h2 id="definition">Definition</h2> <p>反向传播（Backpropagation），即误差反向传播，是一种用于人工神经网络的方法，旨在计算损失函数关于权重的梯度。它通过将误差信号从输出层向输入层反向传播，并利用链式法则更新权重，从而优化模型性能。</p> <h3 id="summary">Summary</h3> <p>一种广泛用于训练神经网络的算法，通过计算梯度来最小化误差。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>梯度下降</li> <li>链式法则</li> <li>损失函数</li> <li>权重更新</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>深度神经网络训练</li> <li>图像分类任务</li> <li>语音识别系统</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/gradient-descent-%E6%A2%AF%E5%BA%A6%E4%B8%8B%E9%99%8D/">Gradient Descent (梯度下降)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/neural-network-%E7%A5%9E%E7%BB%8F%E7%BD%91%E7%BB%9C/">Neural Network (神经网络)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/optimization-%E4%BC%98%E5%8C%96/">Optimization (优化)</a></li> </ul>对齐https://terms-en.ai-term-hub.com/zh/terms/alignment/Sat, 18 Jul 2026 07:43:58 +0000https://terms-en.ai-term-hub.com/zh/terms/alignment/<h2 id="definition">Definition</h2> <p>对齐关注的是确保 AI 系统执行人类真正希望它做的事情，而不仅仅是字面上要求它做的事。它涉及诸如来自人类反馈的强化学习（RLHF）等技术，旨在缩小模型输出与人类期望之间的差距，提高安全性和有用性。</p> <h3 id="summary">Summary</h3> <p>确保 AI 系统的目标和行为与人类价值观及意图相一致的过程。</p> <h2 id="key-concepts">Key Concepts</h2> <ul> <li>RLHF (来自人类反馈的强化学习)</li> <li>价值加载</li> <li>意图匹配</li> <li>奖励建模</li> </ul> <h2 id="use-cases">Use Cases</h2> <ul> <li>调整聊天机器人的礼貌程度</li> <li>确保摘要的事实准确性</li> <li>防止越狱攻击利用</li> </ul> <h2 id="related-terms">Related Terms</h2> <ul> <li><a href="https://terms-en.ai-term-hub.com/en/terms/ai-safety-ai-%E5%AE%89%E5%85%A8/">AI Safety (AI 安全)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/rlhf-%E6%9D%A5%E8%87%AA%E4%BA%BA%E7%B1%BB%E5%8F%8D%E9%A6%88%E7%9A%84%E5%BC%BA%E5%8C%96%E5%AD%A6%E4%B9%A0/">RLHF (来自人类反馈的强化学习)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/ethics-%E4%BC%A6%E7%90%86/">Ethics (伦理)</a></li> <li><a href="https://terms-en.ai-term-hub.com/en/terms/bias-%E5%81%8F%E8%A7%81/">Bias (偏见)</a></li> </ul>