Skip to content

Latest commit

 

History

History
153 lines (120 loc) · 7.65 KB

lesson_1a_course_intro.md

File metadata and controls

153 lines (120 loc) · 7.65 KB
Raw

Lesson 1a: Intro

(30-Oct-2017, live)

Wiki

Wiki: Lesson 1

USF

In-person Info: Deep Learning Certificate Part I

Staff

Classroom

  • 200 students in room at USF
  • 100 Masters' students upstairs at USF
  • 400 International Fellows, via livestream

Being recorded and will become a fastai MOOC.

Teams

  • teams of 6 people
  • get help with stuff

Python

  • using Python 3.6

Deep Learning GPU Platforms

Deep Learning

  • Deep learning is a particular way of doing machine learning
  • Arthur Samuels
    • he invented machine learning
    • rather than programming, step-by-step, give the computer examples
      • let the computer figure out the problem by giving it examples
    • let computer play checkers against itself thousands of times; it figured out which parameters worked the best
    • Samuel Checkers-playing Program appears to be the world's first self-learning program, and as such a very early demonstration of the fundamental concept of artificial intelligence (AI); 1962
    • he worked at Bell Labs and IBM, then Stanford Univ

Machine Learning

Example: ML Algorithm in Predicting Breast Cancer Survival Based on Pathology Slides

  • start with pictures of breast cancer slides
  • work with computer scientists, pathologists worked together to determine features that would predict who would survive or not, based on slides
  • process of building model can take some time (many years); can pass data into ML algorithm, such as logistic regression; regression can determine which sets of features separate out the 2 classes
  • this can work well, but requires a lot of experts and requires the feature data
  • this ML algorithm was more accurate at predicting breast cancer survival than human pathologists

Examples of ML Uses, Thanks to Deep Learning

Future Work

How do we get computers and humans to work better together?

Societal Implications

  • The wonderful and terrifying implications of computers that can learn (Ted Talk by Jeremy Howard 2014)
  • ML / DL algorithms need to be in the hands of practioners who understand the economics / implications of the algorithms.
  • practioners who understand societal implications; what kind of problems should be solved; what does a good solution look like...

Jeremy's Work

Goal of This Course

  • that people from all different backgrounds will use deep learning to solve problems

Deep Learning

  • deep learning is a way of doing machine learning
  • way of giving machine data (examples) and having it figure out the problem that is represented in those examples

What We Are Looking For: Something That Has 3 Properties

3 Things that Give Us Modern Deep Learning

We are looking for a mathematical function that is so flexible that it can solve any given problem.

  1. Infinitely Flexible Functions
  2. All-Purpose Parameter Fitting (way to train the parameters)
  • things can fit hundreds of millions of parameters
  1. Fast and scalable

Example of limitation: linear regression is limited by the fact it can only represent linear functions.

Deep Learning has all 3 of above properties.

  • functional form: neural network
  • multiple layers allows more complex relationships
  • parameters of neural network can be found using gradient descent

Gradient Descent

  • approach works well in practice; local minima are "equivalent" in practice
  • different optimization techniques determine how quickly we can find the way down.

Key discoveries thru Theoretical Side

  • Very, very simple architectures of neural network and very, very simple methods of gradient descent work best in most situations.
  • We'll learn how every step works, using simple math.

Fast and Scalable: Made Possible by GPUs

  • GPU = Graphical Processing Unit
  • GPUs are used in video games
  • Huge industry of video games accidentally built for us what we need to do deep learning
  • GPUs are useful and needed for deep learning
  • GPUs are 10x faster than CPUs
  • Best hardware for deep learning: NVIDIA GTX-1080 Ti for ~ $600

Art of Learning

Projects Done

Work

  • will need to put in 10 hours a week (in addition to lecture time)
  • spend time RUNNING THE CODE (rather than researching the theory)
  • create blog posts

The Test of Whether You Can Understand

  • Deep Learning is about solving problems
    • if you can't turned it into code, you can't solve the problem.
  • You can code / build something with it
  • You can explain / teach it to someone else
    • Write a blog post
    • Help others who have questions

Portfolio

  • people are hired based on their portfolio (not USF DL certificate)
  • GitHub projects, blog posts --> can get hired based on portfolio
  • write down what you are learning in a form that other people can understand

Goal

  • main goal is not to help you move to a deep learning job
  • continue doing what you're doing and bring deep learning to that
  • examples: medicine, journalism, dairy farming
  • opportunities to change society
  • focus: help you be a great practitioner of deep learning
  • opportunity - doing things differently
  • come up with a project idea