DL Workshop 2019

This page includes files and resources related to the 2019 workshop on deep learning, offered 23-24 July, 2019.  Agenda subject to change.

PACR Information Slide: Info Slide

Tutorial material is provided under the GPL license: COPYING


Day 1:

Tutorial 1: Image Processing Essentials

In this tutorial, we present a brief overview of image processing concepts necessary to understand machine learning and deep learning. Completion of this tutorial should give participants the basic background and terminology necessary for an understanding of the mechanics of deep learning.

You can access the Tutorial 1 jupyter notebook here: Tutorial1

You can access the “completed” Tutorial 1 jupyter notebook here: Tutorial1_complete

Images that we will use:
Left click on the image to view full-size, and then right click to download to your computer.

cameraman.png
cameraman.png

peppers.png

peppers.png


Talk 1: From Rules to Machine Learning to Deep Learning

In this talk, we will briefly explore the basic concepts of rule-based learning, classical supervised machine learning, and deep learning.  We will use the MNIST digits dataset as a motivating example in this talk.

You can access a pdf of the slides for Talk 1 here: Talk1


Talk 2: The History of Deep Learning

In this talk, we will provide a brief history of deep learning from the biological perspective, including the perceptron and its biological motivation. 

You can access a pdf of the slides for Talk 2 here:


Tutorial 2: Building a Deep Learning Network for MNIST

In this tutorial, we will introduce the basic structure and common components (convolutional layers, pooling layers, nonlinearities, fully connected layers, etc.) of deep learning networks through a combination of illustrations and hands-on implementation of a network.  While the network is training, this tutorial will be augmented with a talk introducing the basic CNN layers.

You can access the Tutorial 2 jupyter notebook here: Tutorial2

You can access a “complete” Tutorial 2 jupyter notebook here: Tutorial2_complete

You can access a pdf of the slides for Tutorial 2 here: Tutorial2_slides


Day 2:

Tutorial 3: Visualizing the MNIST Network

In this tutorial, we pick up with the same MNIST Network from Tutorial 2 and explore some ways of probing the characteristics of the trained network to help us debug common pitfalls in adapting network architectures.  This tutorial will be augmented with a talk (“That’s not an Ostrich!”) illustrating some humorous “epic fails” of CNNs.

You can access the Tutorial 3 jupyter notebook here: Tutorial3

You can access a “complete” Tutorial 3 jupyter notebook here: Tutorial2_complete

Image that we will use:
Left click on the image to view full-size, and then right click to download to your computer.

my_digits1_compressed.jpg

You can access a pdf of the slides for Tutorial 3 here: Tutorial3 slides


Talk 3: Using the Discovery Supercomputer to Accelerate Training

In this talk we will provide a brief introduction to the use of the Discovery supercomputer at NMSU to help accelerate the training of a network. 

You can access a pdf of the slides for Talk 3 here: Talk3

The code that I ran on Discovery: Discovery_code (on Discovery I typed “sbatch file1.sh” without the quotes to run this code.  Please note–this code points to some directories on Discovery that you will not have access to and thus will not run directly as is.  It is provided instead as an example)


Tutorial 4: The VGG Network, ImageNet Database, and Your Data

In this tutorial, we will explore the use of a common network (VGG-16) and the ImageNet database of natural images.  We will explore common ways in which you can leverage a pre-trained network for other applications.  Participants will be encouraged to use their own data (or additional data recommended for those who didn’t bring data) within the VGG network.  Participants will apply what they learned about network visualization in Tutorial 3 to explore aspects of the VGG network which might be useful for their data and those aspects that might need to be modified.

You can access the Tutorial 4 jupyter notebook here: Tutorial4

Image that we will use:
Left click on the image to view full-size, and then right click to download to your computer.

latest_256_0193.jpg


Talk 4: Beyond Supervised Classification

In this talk, we will provide a brief overview of the applications of deep learning beyond supervised classification.

You can access some materials related to this talk at: Variational_Autoencoders


Please fill out our survey!

Before you leave, please take a few minutes to fill out a brief survey about this workshop.  Your responses are anonymous and will help provide us feedback to improve subsequent offerings of this workshop.

https://survey.nmsu.edu/surveys/?s=3NYJCPECYE