← 返回首页

深度学习

Deep Learning

课程介绍 Course Introduction

学分:4 | 先修课:机器学习导论、线性代数 | 学期:第5学期

深度学习聚焦于多层神经网络模型的理论与实践。课程内容包括前馈神经网络、反向传播算法、卷积神经网络(CNN)、循环神经网络(RNN/LSTM)、Transformer架构、优化算法、正则化技术以及主流框架(PyTorch/TensorFlow)。学生将理解模型设计、训练技巧与表示学习原理,能够应用于图像、语音、自然语言等任务,是AI方向的核心进阶课程。

Deep Learning focuses on the theory and practice of multi-layer neural networks. Topics include feedforward networks, backpropagation, convolutional neural networks (CNNs), recurrent neural networks (RNN/LSTM), Transformers, optimization algorithms, regularization techniques, and mainstream frameworks such as PyTorch and TensorFlow. Students gain skills in model design, training, and representation learning, applicable to vision, speech, and language tasks—a core advanced AI course.

大作业 Final Project

作业标题:图像分类模型

使用深度学习框架实现图像分类模型,完成数据预处理、模型构建、训练优化和评估。对比CNN、ResNet等模型效果,撰写技术报告。

Implement image classification models using deep learning frameworks, complete data preprocessing, model construction, training optimization and evaluation. Compare CNN, ResNet and other models, write technical report.

实施步骤 Implementation Steps

📋 示例:用卷积神经网络做一个图像分类器,比如区分猫狗图片或识别10种常见花卉。你需要用PyTorch搭建ResNet或VGG网络,在公开数据集上训练,调整超参数,最终让测试准确率达到85%以上。
步骤 1
数据集准备
收集图像分类数据集,进行数据增强(翻转、旋转、裁剪等)
产出:增强后数据集
步骤 2
CNN模型设计
设计卷积神经网络结构,参考经典模型(LeNet、AlexNet、ResNet)
产出:模型设计文档
步骤 3
模型实现与训练
使用深度学习框架实现模型,设置损失函数和优化器进行训练
产出:训练好的模型
步骤 4
模型优化
应用正则化、批归一化、学习率调整等技巧提升性能
产出:优化方案
步骤 5
文档提交
整理代码注释和实验报告
产出:最终提交材料

Steps

Step 1
Dataset Preparation
Collect image classification dataset, apply augmentation
Deliverable: Augmented Dataset
Step 2
CNN Model Design
Design CNN architecture, reference LeNet, AlexNet, ResNet
Deliverable: Model Design Doc
Step 3
Implementation & Training
Implement with DL framework, set loss function and optimizer
Deliverable: Trained Model
Step 4
Model Optimization
Apply regularization, batch norm, learning rate scheduling
Deliverable: Optimization Plan
Step 5
Documentation
Prepare code comments and experimental report
Deliverable: Final Submission

Prerequisites · International Exams · Contact · Back to top · Home