Deep Learning Crash Course

Fall 2021

Overview

Deep learning (DL) is currently the most effective way to model the non-structured data, such as images, videos and time series signals. While there are many excellent online resources to teach deep learning with different levels of complexity, it is still of interests to develop and offer a deep learning crash couse for NHLBI DIR community. The motivation behind this course includes:

1. Introduce the basics of deep learning for convolution neural network, recurent neural network and generative model.
2. Introduce practical techniques for model training, debugging, testing, deployment
3. Help everyone be comfort with coding up their models
4. Grow interests inside DIR and improve community awareness of applying DL in biomedical R&D
5. Provide oppertunity for social gathering by bringing colleagues who are interested in these topics together
6. Prepare trainnees and fellows for DL related jobs

Instructors

Hui Xue	David C. Hansen
hui.xue@nih.gov	davidchansen@gradientsoftware.net
Hui is an active researcher on developing deep learning based cardiac imaging applications. The AI imaging products he built are deployed globally and used daily.	David is the founder and technical director of Gradient Software. He developed AI solutions for onocoloy planning, image denoising and enhancement, and reporting.

Assignments

The assignments and lab contents are the meat of this course. The goal is to demonstrate practical skills to solve real-world DL problems using neural networks and to teach useful software packages and tools for effective model development.

After each lecture, there are suggested reading lists. Please read them.

There are 7 assignments in this offering.

Assignments	What it is for
Setup and basic NN	Set up the development environment, including GPU. Code up a NN with MLP and CNN in numpy, for forward-pass, backprop, optimization and training loop etc.
Pytorch model building	Build up the CNN model using pytorch, implement training and validation
CNN for segmentation and detection	Build up the full-size u-net model for segmentation and detection; Resnet; Add your own loss function
RNN for trigger detection	Build up RNN model for trigger detection from time signal; Solve the problem with LSTM, GRU and Transfomer
GAN and fun	Build up the generative adversial network to create new cardaic images; Will your image be good enough to let your model segment it (from assigment 3)?
Data management, experiment manage	Use DVC to manage data and training models; Add wandb for experiment management
Mode deployemnt	Use streamlit to deploy your model from assgiment 2; You can run the model as a web service and show to your lab

After going through the assignments, you should feel comfort/confident to build, debug and deploy your own models! 👏👏👏

Schedule

Time	Lectures	What it is for	Reading list	Assignments
L1	Fundation of deep learning (Hui)	Building blocks, MLP, course overview
L2	Fundation of deep learning, cont. (Hui)	backprop, optimization, loss function ...		Assignment 1
L3	Convolution neural network (David)	Convolution and its variants, badkprop of conv, Common architectures		Assignment 2
L4	Convolution neural network, cont. (David)	Resnet, u-net, use CNN for classification, segmentation, detection		Assignment 3
L5	Recurrent neural network (Hui)	RNN basics, vallina RNN, LSTM
L6	RNN and Transformer (Hui)	GRU, Self-attention, Workd embeding, Transformers		Assignment 4
L7	Model training (David)	Nut and bolts, parallel training, learning rate scheduling, OneCycle, Debug
L8	Generative model (Hui)	Generative vs. discriminative, GAN, its cost function, GAN variants, CycleGAN, CycleGAN+Unet		Assignment 5
L9	Tooling and infrastructures (Hui)	Data management, experiment management, testing, tooling		Assignment 6
L10	Other topics (Hui)	Latest progress, meta-learning, model deployment, build your own DL lab		Assignment 7
L11	Invited lecture 1	Weak supervision and software 2.0
L12	Invited lecture 2	Full stack deep learning