Neural networks and deep learning

This unit introduces the fundamentals of deep learning and its applications across various domains, including image classification, signal processing, and natural language understanding. Neural networks are first described, followed by how training can be achieved with backpropagation. Various forms of deep neural networks are developed, including Multilayer Perceptrons (MLPs), Convolutional Neural Networks (CNNs), and Recurrent Neural Networks (RNNs). Modern advancements such as transformers and Large Language Models (LLMs) are described, as well as their deployment and fine-tuning. The mathematics of optimisation and generalisation is used to interpret and understand the behaviour and training of these networks. Programming frameworks for training, fine-tuning, and deploying neural networks are discussed. Deep learning technologies and design examples are discussed in areas such as visual perception, driverless cars, intelligent assistants, and generative AI.

### 📊 课程难度与压力分析 ECE4179（Neural networks and deep learning）属于 ECE 方向的中高阶课程，学习压力通常来自理论深度、工程实现复杂度和评估标准同时提升。前期可能以概念和小型练习为主，但进入学期中段后会叠加实验、作业、项目和阶段测验，真正挑战在于能否把知识点转化为可验证、可解释的工程结果。建议从第 1 周就建立固定节奏：每周一次概念梳理、一次动手实现、一次结果复盘，把风险前移到平时处理。 ### 🎯 备考重点与高分策略 高分关键是“方法正确 + 过程完整 + 结论可信”。复习建议分三轮：第一轮补基础模型与关键公式，明确适用边界；第二轮按题型做专项训练（建模、推导、仿真、系统设计），形成稳定解题模板；第三轮做限时模拟，训练时间分配与表达清晰度。工程课程评分往往看重“可复现与可解释”，不是只看最后数值。写答案时建议先给框架，再展开关键步骤，最后用工程语言解释结果意义与改进方向。 ### 📚 学习建议与资源推荐 资料优先级可设为：官方讲义与 rubric > lab/tut 指南 > 往年题与补充资源。每周至少安排一次“错误归因”：把问题分成概念误判、推导缺漏、边界条件忽略、实现缺陷、表达不清五类，并记录下次规避动作。对实验/代码任务，保持可复现流程（固定版本、记录参数、保存中间结果）会显著提升提交质量，也能降低期末复习成本。 ### ⚠️ 作业与考试避坑指南 常见失分点包括：只给结论不写关键步骤、忽略单位与量纲、默认理想条件但未说明、实验结论与数据不匹配、代码通过样例但缺边界测试。建议按 D-7 / D-3 / D-1 三阶段推进：D-7 完成主方案与核心验证，D-3 完成边界测试和文档优化，D-1 只做提交核查与口径统一。持续执行 8-10 周稳定闭环后，ECE4179 的成绩通常会更可控，同时会沉淀可迁移到项目和求职的硬技能。