The MATLAB 2023 version of the deep learning toolbox provides a complete tool chain that enables you to model, train, and deploy deep learning in an integrated environment. Compared with Python, MATLAB's syntax is concise and easy to use, without the need for cumbersome configuration and installation, allowing you to implement deep learning tasks faster.
MATLAB's deep learning toolbox provides a wealth of functions and algorithms, covering the entire process from data preprocessing to model training. You can easily import and process large-scale data sets, using batch import and Datastore class functions to perform data operations efficiently. MATLAB also provides an intuitive deep network designer that allows you to quickly build and customize network structures without writing tedious code. At the same time, the collaborative working function of MATLAB and deep learning frameworks such as TensorFlow and PyTorch allows you to flexibly interact with other platforms and give full play to their respective advantages. In addition, MATLAB's deep learning toolbox also has outstanding advantages in model interpretability and feature visualization. You can have a deep understanding of the working principle and decision-making process of the deep learning model through methods such as feature map visualization, convolution kernel visualization, and category activation visualization. MATLAB also provides commonly used interpretability methods such as CAM, LIME, and GRAD-CAM to help you explain and interpret the prediction results of the model. These features will bring deeper insight and understanding to your research and projects.
Chapter One
Introduction to new features of MATLAB 2023a Deep Learning Toolbox
- MATLAB Deep Learning Toolbox Overview
- Introduction and demonstration of Live Script and Interactive Control functions
- Introduction and demonstration of batch big data import and Datastore function functions
- Introduction and demonstration of data cleaning function
- Introduction and demonstration of Deep Network Designer functions
- Experiment Manager function introduction and demonstration
- Introduction to MATLAB Deep Learning Model Hub
- Introduction and demonstration of the collaborative working functions of MATLAB and deep learning frameworks such as TensorFlow and PyTorch
- MATLAB Deep Learning Toolbox Examples简介
Chapter 2, Convolutional Neural Network (CNN)
- The difference and connection between deep learning and traditional machine learning
- The basic principles of convolutional neural networks (What is a convolution kernel? What is the typical topology of CNN? What is the weight sharing mechanism of CNN? What are the features extracted by CNN?)
- The differences and connections between LeNet, AlexNet, Vgg-16/19, GoogLeNet, ResNet and other classic deep neural networks
- Download and installation of pre-trained models (Alexnet, Vgg-16/19, GoogLeNet, ResNet, etc.)
- Optimization algorithms (gradient descent, stochastic gradient descent, mini-batch stochastic gradient descent, momentum method, Adam, etc.)
- Parameter adjustment skills (parameter initialization, data preprocessing, data amplification, batch normalization, hyperparameter optimization, network regularization, etc.)
7. Case explanation: (1) CNN pre-training model to achieve object recognition
(2) Use convolutional neural network to extract abstract features
(3) Customized convolutional neural network topology
(4) 1D CNN model solves the regression fitting prediction problem
8. Practical exercises
Chapter 3, Model Explanation and Feature Visualization
- What is model interpretability? Why do we need to explain CNN models?
- What are the commonly used visualization methods (feature map visualization, convolution kernel visualization, category activation visualization, etc.)?
- Explanation of the principles of CAM (Class Activation Mapping), LIME (Local Interpretable Model-agnostic Explanation), GRAD-CAM and other methods
- Case explanation
Practical exercises
Chapter 4, Transfer Learning Algorithm (Transfer Learning)
1. Basic principles of transfer learning algorithms (Why is transfer learning needed? What is the basic idea of transfer learning?)
2. Transfer learning algorithm based on deep neural network model
3. Case explanation: Model migration based on Alexnet pre-training model
4. Practical exercises
Chapter 5, Recurrent Neural Network and Long Short-Term Memory Neural Network (RNN & LSTM)
1. Basic principles of recurrent neural network (RNN) and long short-term memory neural network (LSTM)
2. The difference and connection between RNN and LSTM
3. Case explanation:
1) Time series forecasting
2) Sequence-sequence classification
4. Practical exercises
Chapter 6, Temporal Convolutional Network (TCN)
1. Basic principles of temporal convolutional network (TCN)
2. The difference and connection between TCN, 1D CNN and LSTM
3. Case explanation:
1) Time series forecast: COVID-19 epidemic forecast
2) Sequence-sequence classification: human action recognition
4. Practical exercises
Chapter 7, Generative Adversarial Network
1. Generative adversarial network GAN (What is an adversarial generative network? Why is an adversarial generative network needed? What can an adversarial generative network do?)
2. Case explanation: automatic generation of sunflower images
3. Practical exercises
Chapter 8 , AutoEncoder
1. The composition and basic working principle of the autoencoder
2. Classic autoencoders (stacked autoencoders, sparse autoencoders, denoising autoencoders, convolutional autoencoders, masked autoencoders, etc.)
3. Case explanation: Image classification based on autoencoder
4. Practical exercises
Chapter 9, Target Detection YOLO Model
1. What is target detection? What are the differences and connections between target detection and target recognition? How the YOLO model works
2. Case explanation: (1) Function introduction and demonstration of the labeling tool Image Labeler
(2) Use pre-trained models to achieve real-time target detection in images, videos, etc.
(3) Train your own data set: Recognition of wearing masks during the COVID-19 epidemic
3. Practical exercises
Chapter 10, U-Net model
1. Introduction to Semantic Segmentation
2. Basic principles of U-Net model
3. Case explanation: Multispectral image semantic segmentation based on U-Net
Chapter 11, Discussion and Q&A
1. How to check literature? (Can you use Google Scholar, Sci-Hub, ResearchGate? Where should you go to find the data and code supporting the paper?)
2. How to refine and explore innovation points? (If it is difficult to do original work at the algorithm level, how can you refine and explore innovative points based on your own actual problems?)
3. Sharing and copying of relevant learning materials (book recommendations, online course recommendations, etc.)
4. Establish a WeChat group to facilitate later discussions and Q&A