Table of Contents generated with DocToc
- Read Online
- Project Description
- Slides
- Course Outline
- Chapter 0 Background Knowledge
- Chapter 1 Introduction
- Chapter 2 Perceptron
- Chapter 3 Multilayer Perceptron (deep feedforward networks)
- Chapter 4 Forward Propagation
- Chapter 5 Learning: Training Neural Networks
- Physical Experiment
- Chapter 6 Make your own neural network to classify handwritten digitals
- Reference
- About Me
This book is powered by Jekyll Book. So you can read it online:
This is my Google Summer of Code 2019 Project with Red Hen Lab.
The Project goal is to design and develops an online course, to teach deep learning for students in the humanities and social sciences. The course will contain basic deep learning theory and labs case studies from multimodal communication.
Project Mentors: Francis Steen, Mark Turner and Rajesh Kasturirangan.
All the course slide can be found at this directory.
- Application of deep learning
- What is Artificial Intelligence?
- What is Machine Learning?
- What is Deep Learning?
- Limitation of deep learning
- Logic-gate neurons
- Neuron-like perceptron
- Neurons are more powerful
- Color Factory
- Multilayer perceptron
- Why is the middle layer called a hidden layer?
- Activation functions
- Commonly used activation functions
- Why must the activation functions be non-linear?
- Forward Propagation
- Example: Handwritten Digits Recognition
- Loss function
- Gradient Descent
- What is Gradient Descent?
- Simple Example
- Avoid Overshooting
- Challenges: Local Minima
- Compute Graph
- Example
- Local Compute
- Compute Graph Advantage
- The Chain Rule
- Compute graph & Chain Rule
- Back propagation
- Back propagation of addition nodes
- Back propagation of multiplication nodes
- Back propagation of ReLU
- Back propagation of Sigmoid
- Application
- Exercise 1
- Exercise 2
- Summary
- A Brief History of CNNs
- Why we need CNNs?
- The Structure of CNNs
- Convolution operation
- Padding Stride
- Convolution
- Pooling
- Some typical CNNs
- Example: Dog or Cat?
- Sequence modeling
- Deep forward neural networks vs Recurrent neural networks
- Recurrent neural networks
- The Problems Of RNNs
- LSTM networks
- Application of RNNs
Programming
Math
- basic matrix, calculus, and statistics.
- What is artificial intelligence, machine learning, deep learning and their relationship?
- Environment Setup Anaconda, TensorFlow and Jupyter Lab.
- How do we learn? (Biological neuron model)
- How can machine learn? (Artificial neural->Perceptron)
- Question and Dataset
- Linear Classifier
- Implement a perceptron
- Nodes
- Input/Output
- Layer
- Input Layer
- Output Layer
- Hidden Layer: Why we call it hidden layer
- Connection
- Fully connected
- Weights
- What is Activation function?
- The common active function
- Regression and classification
- What is Matrix
- Multiplying Matrixs
- Apply Matrix to Neural Network computation
- Design the Output Layer
- How well does the neural network predict: Loss Function
-
Example
-
Loss function: Mean Squared Error
-
Why we use squared error instead of raw error?
The empirical loss measures the total loss over the dataset. Loss function is a function of the Weight.
-
- Gradient Descent
- Minimize error
- What is Gradient Descent?
- Greedy algorithm
- Like Hiking Down a Mountain
- Simple Example
- Local minimum
- Compute Graph
- Chain Rule
- Back Propagation
The error is propagated backwards to the other layers.
- Each student acted as a Neuron
- Mock Forward Propagation
- Mock Backward Propagation
In this chapter, the student will learn how to teach the computer to classify handwritten digits by using MNIST dataset in Python.
DataSet: The dataset I choose for this part is MNIST(Modified National Institute of Standards and Technology) dataset, which has a training set of 60,000 examples, and a test set of 10,000 examples. It is a subset of a larger set available from NIST(National Institute of Standards and Technology) which gives data set of over 800,000 images of handwritten digits from 3,600 writers. The digits have been size-normalized and centered in a fixed-size image.
-
OpenML -- datasets, tasks, flows, results
-
Grokking Deep Learning Andrew W. Trask
-
Introduction to Deep Learning animation videos created by 3Blue1Brown
- Critique of Pure Learning (2019)
- Geoffrey Hinton on capsule networks: https://www.youtube.com/watch?v=rTawFwUvnLE
- Yann LeCun on the limits of deep learning (2016, Quora)
- Artificial Intelligence Pioneer Says We Need to Start Over (2017)
- Why is Geoffrey Hinton suspicious of backpropagation and wants AI to start over?
- Geoffrey Hinton talk "What is wrong with convolutional neural nets?"
- Name: Xinyu You
- Email: [email protected]
- Github: yogayu
- Website: data2art
- Resume: youxinyu.me
If there are any problems, please feel free to contact me.