In the national "14th Five-Year Plan", there are as many as 57 related expressions of "intelligence" and "wisdom". It has become one of the important technical guarantees and core driving forces for my country's "14th Five-Year Plan" period to promote high-quality economic development and build an innovative country. At present, the fiery development of ChatGPT, its basic technology is derived from deep learning. ChatGPT is an artificial intelligence model based on deep learning, and its core technology is neural network. ChatGPT uses a multi-layer neural network to learn and predict the probability distribution of natural language sequences for tasks such as dialogue generation and natural language processing. Deep learning is a machine learning method of multi-layer neural network, which can learn complex features and patterns through training with large amounts of data, and achieve efficient classification and prediction.
In order to actively respond to the needs of scientific research and engineering personnel, according to the "Opinions of the State Council on Promoting the Lifelong Vocational Skills Training System" put forward "keep up with the development and changes of new technologies and new occupations, establish a dynamic adjustment mechanism for occupational classification, and speed up the development of occupational standards" requirements , The Modern Education Research Institute of the Chinese Academy of Management Sciences (http://www.zgyxdjy.com) and Beijing Longteng Asia-Pacific Education Consulting Co., Ltd. jointly held the "Deep Learning Core Technology Practice and Graph Neural Network New Technology Application Seminar" . This training adopts the full combat training mode.
This training is undertaken by Beijing Longteng Asia-Pacific Education Consulting Co., Ltd. and related fees are collected and invoices issued. The specific notice is as follows:
1. Training experts :
Senior experts from the Institute of Automation Technology of the Chinese Academy of Sciences, Beijing Institute of Technology and other scientific research institutions and universities have rich experience in scientific research and engineering technology, and have long been engaged in teaching and research in the fields of artificial intelligence, machine learning, deep learning, and big data analysis. .
- Time and place :
July 27 , 2023 — July 31 , 2023 Shanghai (simultaneously transferred to live broadcast online )
( Registration on the 27th and distribution of class materials , classes from the 28th to the 31st )
3. Training features:
1. Using a simple method, combined with examples and a large number of code exercises, focusing on explaining the deep learning framework model, scientific algorithms, and training process skills.
2. Be able to grasp the technological development trend of deep learning, be able to master the core technology and practical skills of deep learning, and at the same time analyze and explain and discuss the difficult problems in the work, so as to effectively improve the students' ability to solve complex problems;
3. Master the construction and configuration of Transformer training network, a large deep learning model , and master the deep mining of data value.
4. Master the graph neural network model and framework Py Torch
5. Practice cases such as handwritten font recognition and leaf classification, and let AI play games by itself.
6. According to one's own scientific research projects and subject research, master the five framework models of deep learning flexibly.
Note: Other open source public datasets: ImageNet, MS-COCO, UCF101, HMDB51, PASCAL VOC, Open Images, etc.
4. Participants:
All provinces, municipalities and autonomous regions are engaged in artificial intelligence, deep learning, computer vision, face recognition, image processing, pedestrian detection, natural language processing and other related fields. And graduate students and other related personnel, as well as deep learning, computer vision enthusiasts.
5. Fee standard:
Class A: 5680 yuan/person (including registration fee, training fee, materials fee, and class A certificate fee), accommodation can be arranged uniformly, and the expenses are at their own expense.
1. The training fee is collected by the teaching institution that organizes the training class and provides a training invoice.
2. You can enjoy a 10% discount for remittance one week before class, or a 10% discount for registration of more than 5 people, and the two discounts cannot be enjoyed at the same time. Sign up for more than 8 people to enjoy a 12% discount.
3. Students participating in online and offline training can enjoy the rights of video recording and playback, and the right to participate in offline courses of the same theme for free .
6. Issuing certificates:
A. Students who have participated in relevant training and passed the assessment will be issued a "Deep Learning Development and Application Engineer" (Advanced) professional ability certification by the Modern Education Institute of the Chinese Academy of Management Sciences , which can be inquired through the official website, and the certificate can be used by relevant units It is an important basis for appointment, professional title evaluation, professional and technical personnel ability evaluation and assessment.
Note : Students are requested to submit electronic color photos (more than 20KB , red and blue backgrounds are acceptable ) , copies of ID cards and academic certificates to the registration mailbox .
7. Matters needing attention
1. Designated registration email address: [email protected].
2. After the registration is successful, the conference affairs group will issue a specific registration notice and driving route one week before the registration.
3. Students need to bring their own computer, equipped with win10, 64-bit system, 8G or above memory, and reserve 100G of hard disk space.
Attachment: specific course schedule
| key point |
|
|
| 1. Algorithm and scene fusion understanding |
1. Unstructured data with spatial correlation, CNN algorithm . For typical image data, there is a spatial correlation between pixels. For example, image classification, segmentation, and detection are all CNN algorithms. 2. Time - dependent unstructured data, RNN algorithm . A common phenomenon in this type of scenario is that there is a time series correlation between data, that is, there is a sequential dependency between data. For example, natural language processing and speech-related algorithms are based on RNN algorithms . 3. Non- Euclidean data structure , GNN . Such scenarios can typically be represented by graphs. For example social networking etc. |
Case summary explanation Medical field: detection of related diseases such as epidemic diseases and tumors Remote sensing field : such as scene recognition in remote sensing images Petroleum exploration : such as oil particle size detection Rail transit: such as subway dense crowd detection Detection field : such as fault detection Public security field : such as criminal behavior analysis Defense field : target detection, signal analysis, situational awareness... Economic fields: such as stock forecasting |
| 2. Data understanding and processing |
Analyze typical data in typical scenarios, and process the data in combination with specific algorithms |
1. Structured data, how to read and organize the data. 2. Image data, processing methods in the actual application process, how to do data preprocessing, data enhancement, etc. 3. Timing signals, how to combine single-point data into a sequence, and the basic method of sequence data processing. |
| 3. Technical path design |
Design specific neural network models for specific scenarios, and introduce typical data-adapted network structures. |
1. Basic principles of DNN model building 2. Common network structure and parameter analysis in CNN model. 3. Some basic operators supported in RNN, how to organize sequence data. |
| 4. Model validation and troubleshooting |
Simple algorithms or models quickly verify typical scenarios, and explain some frequently occurring problems. |
|
| 5. Principles of advanced model optimization |
Different models need to use the optimization function and the optimization method of parameters in backpropagation |
1. Introduction to the algorithm of model optimization, and the introduction of the algorithm based on stochastic gradient descent. 2. Introduction of loss functions adapted to different scenarios. 3. The pushing process of the backpropagation gradient for typical scenarios. |
| 6. Advanced-Customization Ideas |
Combined with some projects of previous students, briefly introduce the idea of solving a specific problem. |
In remote sensing imaging, the identification of crop types in plots. |
| Practical analysis and training The first stage: 神经网络实践 |
实验：神经网络 1.神经网络中基本概念理解：epoch、batch size、学习率、正则、噪声、激活函数等。 2.不同的数据生成模型、调整网络参数、调整网络规模 3.神经网络分类问题 4.不同数据特征的作用分析、隐含层神经元数目 5.过拟合 高频问题： 1.输入数据与数据特征 2.模型设计的过程中的参数与功能的关系。 关键点： 1.掌握神经网络的基本概念 2.学会搭建简单的神经网络结构 3.理解神经网络参数 |
|
| 实操解析与训练 第二阶段： 深度学习三种编程思想 |
实验：Keras实践 1.理解Keras基本原理 2.学会Keras编程思想 3.三种不同的深度神经网络构建编程方式 4.给定数据集，采用Keras独立完成实际的工程项目 高频问题： 1.如何编程实现深度神经网络 2.三种开发方式的具体使用 关键点： 1.掌握Keras编程思想 2.采用三种不同方式编写深度神经网络 |
|
| 实操解析与训练 第三阶段：CNN实践 |
实验：图像分类 1.使用CNN解决图像分类问题 2.搭建AlexNet 3.VGG16/19 4.GoogleNet 5.ResNet 高频问题： 1.CNN更复杂的模型在哪里可以找到代码 关键点： 1.使用卷积神经网络做图像分类 2.常见开源代码以及适用的问题 |
|
| 实验：视频人物行为识别 1.基于C3D的视频行为识别方法 2.基于LSTM的视频行为识别方法 3.基于Attention的视频行为识别方法 高频问题： 1.2D卷积与3D卷积 2.视频的时空特征 关键点： 1.C3D网络的构建 2.Attention机制 |
||
| 实操解析与训练 第四阶段： R-CNN及YOLO实践 |
实验：目标检测 1.目标检测发展现状及代表性方法 2.两阶段目标检测方法：R-CNN系列模型 3.一阶段目标检测方法：YOLO系列模型 高频问题： 1.提名与分类 2.BBOX实现策略 3.YOLO Loss函数 关键点： 1.提名方法 2.ROI Pooling 3.SPP Net 4.RPN 5.YOLO |
|
| 实操解析与训练 第五阶段： RNN实践 |
实验：股票预测 1.股票数据分析 2.同步预测 3.异步预测 高频问题： 1.历史数据的使用 关键点： 1.构建RNN 2.采用Keras编程实现 |
|
| 实操解析与训练 第六阶段： Encoder-Decoder实践 |
实验：去噪分析 1.自编码器 2.去噪自编码器 高频问题： 1.噪声的引入与去除 关键点： 1.设计去噪自编码器 |
|
| 实验：图像标题生成 结合计算机视觉和机器翻译的最新进展，利用深度神经网络生成真实的图像标题。 1.掌握Encoder-Decoder结构 2.学会Seq2seq结构 3.图像CNN +文本RNN 4.图像标题生成模型 高频问题： 1.如何能够根据图像生成文本？ 关键点： 1.提取图像特征CNN，生成文本RNN 2.构建Encoder-Decoder结构 |
||
| 实操解析与训练 第七阶段： GAN实践 |
实验：艺术家作品生成 1. 生成对抗网络原理 2.GAN的生成模型、判别模型的设计 高频问题： 1.生成模型与判别模型的博弈过程 关键点： 1.掌握GAN的思想与原理 2.根据需求学会设计生成模型与判别模型 |
|
| 实操解析与训练 第八阶段： 强化学习实践 |
实验：游戏分析 1.游戏场景分析 2.强化学习的要素分析 3.深度强化学习 高频问题： 1.DNN 与DQN 2.探索与利用 关键点： 1.深度强化学习的原理 2.根据实际需求，设计深度强化学习模型 |
|
| 实操解析与训练 第九阶段： 图卷积神经网络实践 |
实验：社交网络分析 1.图神经网络的原理 2.图卷积神经网络的思想 3.设计图卷积神经网络进行社交网络分析 高频问题： 1.如何从图神经网络的原理转化到实际编程 关键点： 1. 掌握图神经网络原理 2. 图卷积神经网络编程实现 |
|
| 实操解析与训练 第十阶段： Transformer实践 |
实验：基于Transformer的对话生成 1. Transformer原理 2. 基于Transformer的对话生成 3.基于 Transformer 的应用 高频问题： 1.如何应用自注意力机制 2.如何应用于自然语言处理与计算机视觉 关键点： 1.self-Attention机制 2.position |
|
更多资讯请关注微信公众号：人工智能技术与咨询或登录中国人工智能培训网