How to Get Started with Machine Learning

class: center, middle

# Tools to Get Started with Machine Learning / AI
## Christian Hudon | JDA Labs

.footnote[Twitter: @christian_hudon and @JDALabsMTL]

---
# You think machine learning is cool

.middle[
<video height="400" autoplay loop alt="Not Hotdog, as seen on HBO">
  <source src="img/Not_Hotdog.webm" type="video/webm">
</video>
]

---
# You've taken a first online class

![](img/Coursera_Deep.png)

---
class: center

# But what about after that?

## Maybe you have a first project idea...

## (Or maybe not.)

???
* That's cool. We'll also have some suggestions for good first projects in here later.

---
class: center, middle

# Let's Get Started!

???
# First, pick a deep learning library

---
class: center, middle

![](img/libraries_word_mural.png)

???
Well. ok. Let's put that off a bit, and start by visualizing our data instead.

Amazon: MXNet
Baidu: PaddlePaddle
Facebook: PyTorch
Google: TensorFlow
Microsoft: CNTK / Microsoft Cognitive Toolkit
Tencent: FeatherCNN

Collage of the names of: theano, lasagne, nolearn, Keras, Pylearn2, Blocks, Caffe, Mxnet, CNTK, Torch7, Pytorch, Sklearn-theano.

---
.center[![Python Visualization Landscape](img/python_viz_landscape.png)]

???
* Worse.
* Easier to pick once experienced, but if you're starting out...

---
class: center, middle

# Who is this presentation for?

???
* No ML experience necesary, but ML not taught here
* A little bit of programming experience
* Not afraid to get feet wet

* Gets you started. Initial selection of tools and starter project suggestions
* Beginner-friendly tools (not necessarily what we use here)
* Python (that's what I know best / deep learning research happening in Python)

Giving you a good starting point to continue learning on your own.

---
class: center, middle

# Trying things out:

# Jupyter Lab

???
* Born out of reproductible science movement
* Runnable document
* Code, text, plots, equations in same document
* Lab: big brother, ready for use (stable plugin API: 1.0)

---
class: center, middle

# Numerical computations:

# Numpy

???
* The standard foundation in Python for numerical work
* Other libarries either built on top, or work with Numpy.
* Similar to Matlab...

---
class: center, middle

# Data Exploration:

# Pandas

???
* Data cleaning and exploration a significant part of the work!
* Pandas
 * Easy to use yet powerful
 * Very good support for dates, categoricals

---
class: center, middle

# Classical machine learning:

# Scikit-learn
## For NLP: SpaCy

???
* Easy to use, models know how to train themselves
* Large number of algorithms
* Excellent documentation
* Ability to work with sequence of preprocessings + model as a composite model (pipelines)

---
class: center, middle

# Deep learning:

# Pytorch (and Skorch)

.footnote[.red[*] Keras (with Tensorflow) also a reasonable choice.]

???
* Created by Facebook; rapid growth in use
* Fairly often described as "fun"
* Easier to understand causes of errors
* Skorch: makes it easy to replace scikit-learn model
* Keras and TensorFlow: harder to debug problems, but new "immediate" mode helps

---
class: center, middle

# Visualization (simpler / pre-canned):

# Pandas & Seaborn

???
* Quick and easy!
* For the basics

---
class: center, middle

# Visualization (expressive):

# HoloViews (or Altair)

???
* Easy to try complex plots quickly
 * One of the good ideas of R
 * Difference between 1-line plots and 1 paragraph ones!
* Great for visually exploring your data
* HoloViews:
 * Interactive plots in browser
 * Scalable to huge datasets with `datashader`

Be picky about your visualization library! Invest the time!

---
class: center, middle

# How to continue learning?

## Do a project!

---

# Dataset Sources

* The Open Data movement (Bixi, etc.)

* Kaggle

* Your own life! (FitBit, your bank)

* The Internet!

---

# Project Ideas

* Forecast Bixi usage

* Cat vs. dog detector

* Automated categorization of credit card bill items

???

* Reinforcement learning not a good first project.
* Landing near the top of the Kaggle leaderboard needs different skills.
* Creating trading models that make money has gotchas.

---

# After your first project

* Getting better at Python: https://www.youtube.com/playlist?list=PLu5kChXP9Hipqr1gIROxGlvb3YVadOjUB (14 videos currently)

* Learning the basics of classes and object-oriented programming: http://www.greenteapress.com/thinkpython/html/thinkpython016.html (and next two chapters)

* The matrix algebra you need for machine learning: http://parrt.cs.usfca.edu/doc/matrix-calculus/

* Scaling with [Dask](https://dask.pydata.org/en/latest/) and [Dask-ml](https://dask-ml.readthedocs.io/en/latest/) and [Dask-Kubernetes](https://dask-kubernetes.readthedocs.io/en/latest/)

* Learning the basics of SQL

---

.pull-left[
## Thank you!

Slides and notebooks: http://christianhudon.name/talks/#DSDT-2018-06
]

.pull-right[
## Tools

* Trying things out: **Jupyter Lab**

* Numerical computations: **Numpy**

* Data Exploration: **Pandas**

* Classical machine learning: **Scikit-learn** (and SpaCy)

* Deep learning: **Pytorch** (and Skorch)

* Visualization: **Pandas** and **Seaborn**

* More Visualization: **HoloViews** (or Altair)
]