How to learn AI efficiently

My understanding of AI learning: know what it is, know what it is, don't ask for a deep understanding, and do it again. It can also be understood as the three processes of AI learning.

Know it: Know the models and methods of AI, how to use these models and methods to solve problems, understand the models and some frameworks. For example, image classification, we know that there are commonly used networks such as VGG, ResNet, EfficientNet, MobileNet, etc., which can be used to solve image classification problems. In short, use it first.

Know why: know the implementation of AI models and methods, learn some classic papers, and grasp the development trend of this industry. Taking image classification as an example, we know the structure of ResNet's Block, the structure of the entire network, and the methods used in the author's paper to improve the performance of the algorithm. These are helpful for us to build our own models in the future.

Don't ask for deep answers: Don't worry about the interpretability of the AI ​​model, and don't worry about the meaning of each layer word by word, or even the setting of parameters. At present, AI has no fixed theories and theorems, and many methods and models are also experimental. For most people, these are meaningless.

Hands-on and hands-on: Li Mu wrote a book called "Learning Deep Learning by Hand". The name of the book has already told us how to learn deep learning? That is hands-on. Unfortunately, the reality is that many people speak the models and methods well and can even memorize the papers, but they cannot write the simplest programs, such as image classification and text classification. Putting a lot of time in theory, the hands-on ability is almost zero, some can't write python, others can't understand what they write, some can't load data, and some can't call models. All in all, it is a lack of hands-on ability.

Learning Path

I have summarized several learning paths:

1. Take AI-related courses. If you are a student, you can participate in the school's courses, and if you are already working, you can watch some public courses. This way of learning knowledge is more theoretical, after all, you can't keep talking about code in class.

2. Buy books. AI books are theoretical, such as "Deep Learning" and "Statistical Methods", which are more suitable for learning theoretical knowledge. If you are a beginner, it is very difficult to learn and it is difficult to persevere. More of them are more practical. These books have a common feature that the examples are basically the same, and the code in the book is relatively old, and some modifications are required to run the code.

3. Pay attention to some official accounts and blogs. The updated content of official accounts and blogs is still relatively new, and it can also help us solve some problems, but there are also shortcomings, that is, lack of systematicness, and the code used in many articles has not been verified by the author. . Caused us to browse a lot of articles and did not solve the problem.

4. Participate in some training. Participating in training is a way to learn quickly. The speed mentioned here is to compare with the above. For office workers, job promotion and career change are good choices. There is a saying, free is often the most expensive, I think it is more expensive in terms of time. It is also a good choice if an organization can summarize the knowledge, achieve situational teaching, and practical training, so that students can quickly grasp the knowledge.

I found a good training tutorial on the Internet and shared it with you.

Notice of the China Academy of Management Science on Holding the "Transfer Learning Core Technology Development and Application Seminar" (qq.com)

Next, let's take a look at the course together.

Target

1. In-depth understanding of the composition, training and implementation of neural networks, and master key concepts such as deep spatial feature distribution;

2. Master the ideas and basic forms of transfer learning, understand the basic methods of traditional transfer learning, and compare the advantages and disadvantages of various methods;

3. Master the ideas and components of deep transfer learning, and learn various methods of deep transfer learning;

4. Master the cutting-edge methods of network structure design and objective function design of deep transfer learning, and understand the application of transfer learning on PDA and Source-Free DA;

5. Master the application of deep transfer learning in tasks such as semantic segmentation, object detection, and pedestrian re-identification, learn image/video grid transfer methods, and understand the application of grid transfer in real life;

6. Understand and master cutting-edge methods such as small sample learning and Transformer, as well as basic ideas and applications in practical scenarios;

7. Master image and video format migration, cross-domain semantic segmentation in autonomous driving, and target detection through practical operation.

For people

Mainly for the staff of enterprises and institutions.

Specific course schedule

Basic knowledge

1. Introduction to Machine Learning and Introduction to Classic Machine Learning Algorithms

content

1. What is machine learning?
2. Machine Learning Framework and Basic Components
3. The training steps of machine learning
4. Classification of Machine Learning Problems
5. Introduction to Classic Machine Learning Algorithms

Chapter Objectives

Machine learning is one of the important technologies of artificial intelligence. A detailed understanding of the principles, mechanisms and methods of machine learning will lay a solid foundation for deep learning and transfer learning.

2. Introduction to Deep Learning and Introduction to Classical Network Structure

content

1. Introduction to Neural Networks
2. Introduction to Neural Network Components
3. neural network training method
4. Introduction to Convolutional Neural Networks
5. Introduction to Classic Network Structure

Chapter Objectives

In-depth understanding of the composition, training and implementation of neural networks, mastering key concepts such as deep spatial feature distribution, and laying a knowledge foundation for deep transfer learning

3. Fundamentals of Transfer Learning

content

1. Introduction to Transfer Learning
2. Sample-Based Transfer Learning
3. Feature-Based Transfer Learning
4. Transfer Learning Based on Classifier Adaptation

Chapter Objectives

Master the ideas and basic forms of transfer learning, understand the basic methods of traditional transfer learning, compare the advantages and disadvantages of various methods, and grasp the scope of application of transfer learning.

4. Introduction to Deep Transfer Learning

content

1. Introduction to Transfer Learning
2. Sample-Based Transfer Learning
3. Feature-Based Transfer Learning
4. Transfer Learning Based on Classifier Adaptation

Chapter Objectives

Master the ideas and components of deep transfer learning, learn various methods of deep transfer learning, compare the advantages and disadvantages of various methods, and master the scope of application of deep transfer learning.

Transfer learning methods and applications

V. Introduction to Frontier Methods of Transfer Learning

content

1. Deep transfer network structure design
2. Deep transfer learning objective function design
3. Transfer Learning in New Scenarios

Chapter Objectives

Master the network structure design of deep transfer learning, the cutting-edge method of objective function design, and understand the application of transfer learning on PDA and Source-Free DA.

6. Frontier applications of transfer learning

content

1. Application of Transfer Learning in Semantic Segmentation
2. Application of Transfer Learning in Object Detection
3. Application of Transfer Learning in Person Re-ID
4. Image and video style transfer

Chapter Objectives

Master the application of deep transfer learning in tasks such as semantic segmentation, object detection, and person re-identification, learn image/video grid transfer methods, and understand the application of grid transfer in real life.

7. Cutting-edge methods and applications such as small sample learning and Transformer

content

1. Introduction to small sample learning concepts and basic methods
2. Small sample learning application
3. Transformer concept and basic method introduction
4. The application of Transformer in the image field

Chapter Objectives

Master the cutting-edge methods and basic ideas of small sample learning and Transformer, and understand the application of small sample learning and Transformer in practical scenarios.

Environment construction

1. Hardware preparation: GPU memory more than 11GB

2. Software preparation: Linux operating system (Ubuntu16.04 or above), graphics card driver installation (512.54), CUDA Toolkit (10.1) and cuDNN acceleration library (7.6.4), VS Code editor installation, Jupyter Notebook

3. Programming language and framework: Python3.8.5, torch1…07、torchvision0.8.2、mmcvfull1.3.7、opencv-python4.4.0 、 matplotlib3.4.2、numpy1.19.2、Pillow8.3.1、scikit-learn1.0.2

4. Data set preparation: Office-31, IRVI, GTA5, Cityscapes, Foggy cityscapes, etc.
   Note: The organizer provides cloud server for hardware preparation

5. 

Deep Transfer Learning Practice

1. Master the basic principles and programming ideas in PyTorch.

2. Understand when and how to perform transfer learning in a new scenario or dataset.

3. Load data, build models, train networks, and fine-tune networks with PyTorch.

4. Given a migration scenario, domain adaptation in image classification is independently accomplished using the daib library and generative adversarial techniques.

5. To visualize the migration effect, use t-SNE in the machine learning library scikitlearn to visualize the migrated high-dimensional data.

   

Image and Video Format Migration Practice

1. Master the grid migration technology based on Generative Adversarial Networks.
2. The construction of image/video format transfer network, focusing on mastering the internal logic of encoder and decoder and the use of different loss functions.
3. Practice infrared video conversion to visible video format migration

Cross-domain semantic segmentation practice in autonomous driving

1. Master the development status and representative work of semantic segmentation, such as FCN, DeepLab series, etc.

2. Learn about common semantic segmentation evaluation metrics (PA, mPA, mIoU, FWIoU) and common datasets (PASCAL VOC2012, ADE20K, BDD100K, Cityscapes, GTA5, Dark Zurich).

3. The understanding and use of the semantic segmentation toolbox MMSegmentaion.

4. Designing a segmentation model can migrate from data obtained in a simulated environment to data generated in a real-world scenario.

   

Object Detection Practice

1. Master the basic framework of target detection algorithms and classic models in target detection, such as the two-stage detection model of the R-CNN series and the single-stage detection model of the YOLO series.
2. Master the evaluation metrics (IOU and mAP) of object detection models, standard evaluation datasets (Pascal VOC, MS COCO, and Cityscapes), and some training techniques in detection models, such as data augmentation, multi-scale training/testing, and fine-tuning of prediction boxes /voting method, online hard case mining, softening non-maximal suppression, RoI matching and integration.
3. Practice the construction of an end-to-end object detection framework based on Transformer, and compare the migration performance with the CNN-based network on the new dataset.
4. 

Training experts

The training experts are senior experts from the Institute of Computing Technology of the Chinese Academy of Sciences, Tsinghua University, Beijing Institute of Technology and other scientific research institutions and universities. They have rich
experience , and have been engaged in artificial intelligence chips for a long time. , FPGA, deep learning and other fields of teaching and research work.

Students who pass the assessment will be issued the "Transfer Learning Core Technology Development and Application Engineer" (senior) post certification by the Institute of Modern Education, China Academy of Management Sciences, which can be
inquired through the official website (http:// www.zgyxdjy.com), this certificate can be used as an important basis for the ability evaluation, assessment, appointment and professional title evaluation of professional and technical personnel in relevant units.

For those who need a certificate, the study is over, but for those who want to study AI, it is just the beginning. I think that after training you can master the dots, and you need to continue to expand and connect the dots and build your own body of knowledge to be perfect.

Sun Gong's death is endless, and Tang donations will end up in the sea! persist in!