Deep learning Part 2 Vision Algorithms

This class is part two in a series that takes engineers from zero to one in deep learning.  It’s designed as a follow up to Technical Introduction to AI, Machine Learning & Deep Learning but could also be appropriate for someone who had done some machine learning and wanted to really focus on deep learning algorithms.

This is a hands-on course that takes students from little knowledge of deep learning to comfort building real world vision models.  It requires very little math, but reasonably proficient programming skills.  At the end of class students will be able to build, inspect debug and deploy deep learning for a variety of applications in vision.  

Technologies Introduced and Used:

• Deep Learning/Data Science Frameworks

• numpy

• keras

• tensorflow

• Deep Learning Model Architectures

• Multi-Layer Perceptron

• Convolutional Neural Networks

• Adversarial Networks

• Applications

• Object detection

• Image segmentation

• Bounding Boxes

Prerequisites:

This is a sequel to our first-course Introduction to AI, Machine learning & Deep Learning.  It is a series designed for practicing engineers who want to get into deep learning.

You can also skip the first course if you have experience with machine learning and feel like you know most of what is covered in part one.  

The entire course is done in python, so if you are unfamiliar with python you should brush up.  No math is required but we will use a little bit of calculus.  It’s ok if you don’t follow along in those parts.

What you need to bring:

Students need to bring a laptop.  We have detailed setup instructions at https://github.com/lukas/ml-class/blob/master/README.md

If you want to run on a GPU either on your laptop or in the cloud for this course, you can.  We have setup instructions for an AWS machine at https://github.com/lukas/ml-class/blob/master/aws.md

Take aways:

-Practical high-level knowledge of how deep learning algorithms actually work

-How to install the frameworks so they take advantage of your GPUs

-How to build models from scratch

-How to debug models when they don’t work

-How to fine tune popular models like Inception and ResNet when training data is limited

-How to deploy models effectively

Curriculum:

Morning: Introduction to Neural Nets

9:00 – 10:00 Overview of deep learning history and terminology and why it matters.  Loss functions, backpropagation and more.

10:00 - 12:00 Overview of keras, tensorflow and numpy and how they are all different.  Build a small neural network from scratch together in python to do digit recognition.

12:00-1:00 Lunch

Afternoon: Vision

1:00 - 2:00 Transfer learning and fine-tuning.  Build a dog vs cat classifier in several different ways.

2:00 - 3:00 Build and deploy and improve a facial emotion classifier.

4:00 - 5:00 How to apply the work to semantic segmentation and bounding box detection.  Adversarial learning and the future of vision classifiers.

Testimonials and Feedback:

"I found it to be really engaging and interesting. I was already familiar with some ML concepts, so it helped me understand them better and think about how to apply them. The code samples are really great and will definitely reference them in the future. I thought the class went at a generally good pace."

"Good experience - full of great resources and discussion. Good, practical intro for new folks, and also valuable for those familiar with the basics. I walked away excited to experiment!"

“Class was great, you ticked off my curiosity. I am excited to review the content and retry it by myself. Thank you for encouraging peer to peer collaboration and making the effort to build the slack channel. I think it was nice to see you debug live.”