This repository contains the complementary material for 项目实践 and 机器学习基础 taught at SES. Due to the distributed and asynchronous nature of this study format, we will steer the course through a series of self-contained units around programming, visualization, and simulation, which may be your interest for your future endeavor. This course tries to fulfill this arrangement by introducing some tools and demonstrating their capability.
A quck list of topics:
Here is the slide:
pdf for introduction to the smart car
题目：介观神经科学研究的知识图谱系统开发
With the progress of the whole brain mesoscopic scale imaging system, such as MOST, studies on the brain connectome go to the next generation focusing on the finer brain structures. However, previous research results about individual brain structure are often scattered around a large set of literature, it prohibits the researchers to catch up with the state-of-art research. In this project, you will develop an end-to-end, ready-to-use search system to allow your users to search the mesoscopic scale neuroscience terms and get back with the relevant literature, just as the general search engine. The corpus and the specific trained Bidirectional Encoder Representations from Transformers (BERT) model are ready and waiting for you to present them to broader audience.
You may consider to join if you are interested/proficient in web development, natural language processing, and knowledge graph.
A team of 3 or 4 students is preferred for this project.
Project time span is expected to be 4 months part time.
题目：基于Rust的高效GPU内存分配系统开发
Deep learning is in its period of prosperity and the process of producing the production-ready models have evolved into the front end and the back end, where the specialists construct the model for a specific task during the front end phrase and the devleopers fine tune the model for the deployment. In this project, you are invited to extend an existing Rust-based deep learning library with a GPU memory allocator to allow it to run efficiently on the GPU side and you will get yourself familar with the high performance GPU programming. Rust is a segfault-free C-like-speed programming language even with a naive implementation.
You may consider to join if you are interested/proficient in Rust, deep learning, and GPU programming.
A single person is preferred for this project.
Project time span is expected to be 8 months part time.
题目：用于多媒体内容原型生成的文图转换方法研究
Text and Image often come together in the multimedia content supply. There are already amazing researches on the image-to-text conversion (image captioning). However, the text-to-image conversion is relative challenging as the information is usually extended in this case. You are invited to join this project to develop a prototype system for the text-to-image conversion. The plan starts with the existing public baseline model and develops a model for a specific application scenario.
You may consider to join if you are interested/proficient in deep learning, natural language processing, and computer vision.
A team of 6 is preferred for this project.
Project time span is expected to be 12 months part time.
Electrical motors are used everywhere to drive the world forward. In this section, we will try to build a motor and a driver for it.
There are several types of electrical motors and different motor needs a different driver circuit to supply the power they run on. There are AC motors, BLDC motor, and other types of motor, we will first start to drive DC motors with brush.
The introduction for the DC motor driver project is presented here:
pdf for introduction to the brush DC motor driver
The videos that show how to use KiCAD to produce a PCB board are listed here:
- Plot the schematic design: @acfun, @bilibili,
- Assign the footprint for the component: @acfun, @bilibili.
- Layout the PCB: @acfun, @bilibili.
Please do you research and find proper components to build up the PCB board for a 4-channel DC driver.
4 units are planned for the first 4 weeks. Each unit will contain at least 2 parts. The 1st part will discuss general practice and methodology around software engineering and this serves as the general topic across the board. The second part will introduce some tools or libraries and demonstrate how to use them. It’s highly recommended that you should do your own research about this tool or library before the class and get familiar with it. The class time will be the Q/A time and I will stay online to answer your question through Github issues, QQ, and Wechat.
In the introduction to software engneering, we will cover the concept of software, software engineering, and software crisis. We also show the basic workflow of Git, the distributed version control system.
We will have 2 parts in this unit. In the first part, we go through an agile method called Scrum. In this second part, we will talk about Python.
We will introduce the object oriented design and programming pattern in this unit. A midterm project is also in this unit.
This unit is about software verification and validation method, focusing on unit test.
Please form a team with 2~3 members and pick a project from http://www.cse.msu.edu/~cse231/PracticeOfComputingUsingPython/.
Q: Do we have a syllabus for this course? What's the goal of this course?
A: For the situation caused by the Coronavirus, we do have an alternative plan for the course for the first 4 weeks. The topics are listed above. The goal of this course is to introduce several concept and practise around software development.
Q: The videos can be more connected and the title of the video is not clear?
A: We want each video to be self-contained and independent as possible so that referece to these video in the future makes more sense without refering to the context and the background. Acfun.cn and bilibili.com are meant to be used as video host only. Title and description for the video are only available on the site where this readme resides. This makes the material for this course neutral to video hosting site.
Q: There are some parts where complementary slides are not available?
A: The programming parts are mainly screen casting.
Q: Team up may be a better choice?
A: Exercise of some units will be for individual, while other exercises require a teamwork.
Q: What about students with a computer at his/her disposal?
A: We provide reference in each unit and some reading instruction. This is for students without computers.
- Machine learning and related terms.
- Types of problem discussed in machine learning.
- Introduction to neural networks.
video: introduction to machine learning @acfun
video: introduction to machine learning @bilibili
The material is in c1_intro.
- logistic regression
- stochastic gradient descent
- input normalization
- oversampling/undersampling
video for logistic regression, @acfun
video for logistic regression, @bilibili
The task in this exercise is for you to apply logistic regression on ``Breast Cancer Wisconsin (Diagnostic) Data Set''.
- You can set train/test split to 0.7/0.3 or any value you like.
- You can choose your favorite tool besides PyTorch.
- Post your answer in a pull request to this repository.
- Discuss with each other while try to work the code out by yourself.
There is a piece of code demonstrating logistic regression using PyTorch. mlex1/lr.py
The data is available at https://archive.ics.uci.edu/ml/datasets/Breast+Cancer+Wisconsin+(Diagnostic)
The data is also available on Kaggle: https://www.kaggle.com/uciml/breast-cancer-wisconsin-data
Note:
- UC Irvine Machine Learning Repository is the classic goto site to find some machine learning dataset.
- Kaggle is a popular machine learning competition site.
- Ensamble models by bagging/stacking/boosting
- MLP by stacking logistic regression
- Auto difference i.e. back propagation
Videos:
- multiLayer-perceptron @acfun
- multiLayer-perceptron @bilibili
- back propagation @acfun
- back propagation @bilibili
pdfs:
- Perceptron and multiLayer-perceptron
- Back propagation slides from Stanford cs231n
- Back propagation slides from Stanford cs231n (source)
- Convolutional operator in convolutional neural network (CNN)
- Other layers often used in CNN
- Learning tricks
Videos:
pdf:
Videos:
pdf:
Please go through deep learning with PyTorch: a 60 minute blitz by yourself. You can post any question in the QQ chat group if you'd like to discuss.
The stock managment system development is the main focus for 项目实践 in 2020 fall. The example repository can be found in ./stock.