The document introduces competitive machine learning and provides an overview of what it is, why people compete, and where competitions can be found. Competitive machine learning involves using machine learning techniques to solve problems by analyzing messy and complex real-world data from sources like Netflix and Kaggle to build models that are evaluated on public leaderboards. Competitions help people gain experience and learn new techniques while allowing them to build portfolios and network, with some competitions offering prizes and job opportunities. Popular past competitions include the Netflix Prize and ImageNet challenges. The document recommends the Titanic competition on Kaggle as a good starting point for beginners.

1. Introduction to
Competitive
Machine
Learning
Matt Motoki & Thomas Yokota

2. Overview
● What is Competitive Machine Learning?
● Why Compete?
● Where Can I Compete?
○ Netflix
○ ImageNet
○ Kaggle
● How Do I Start?
○ Kaggle Walkthrough
○ Titanic Competition

3. What is Competitive
Machine Learning?

4. What Is Competitive Machine Learning?

5. What Is Competitive Machine Learning?
1. Problem
a. Business
b. Research
2. Data
a. Messy
b. Disparate sources
c. Different formats (text, tabulated, images, etc.)
3. Crowd-sourced
a. Teaming up
4. Knowledge & Tools
a. Python, R, Keras, TF, C++, etc.
5. Models
a. Benchmarked (evaluation score)
b. Leaderboard

6. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
● Customer Interaction
● In-depth theory
Problems with Competitions
● Gains are incremental
● Addictive

7. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
● Customer Interaction
● In-depth theory
Problems with Competitions
● Gains are incremental
● Addictive

8. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
Problem

9. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
● Customer Interaction

10. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
● Customer Interaction
● In-depth theory

11. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
● Customer Interaction
● In-depth theory
Problems with Competitions
● Gains are incremental

12. Downsides to Competitive Machine Learning
What Competitive Machine Learning Lacks
● Problem Formulation
● Customer Interaction
● In-depth theory
Problems with Competitions
● Gains are incremental
● Addictive

13. Why Compete?

14. Why compete?
● Gain experience

15. Why compete?
● Gain experience
● Build a portfolio

16. Why compete?
● Gain experience
● Build a portfolio
● Learn new techniques
Inception Network

17. Why compete?
● Gain experience
● Build a portfolio
● Learn new techniques
● Networking

18. Why compete?
● Gain experience
● Build a portfolio
● Learn new techniques
● Networking
● Prizes

19. Why compete?
● Gain experience
● Build a portfolio
● Learn new techniques
● Networking
● Prizes
● Job opportunities

20. Where Can I Compete?

21. Netflix Prize

22. Background
● Improve Netflix’s Cinematch algorithm by 10% (RMSE).
● 100 million movie ratings, 480,189 movies and 17,770 users
● Oct 2006 – Sep 2009
● $1,000,000 prize

23. Background
● Improve Netflix’s Cinematch algorithm by 10% (RMSE).
● 100 million movie ratings, 480,189 movies and 17,770 users
● Oct 2006 – Sep 2009
● $1,000,000 prize
Winning Solution
● Matrix factorization
● Neural Networks
● Gradient Boosted Trees

24. ImageNet
Large Scale Visual
Recognition Challenge

25. Background
● Started in 2010
● 14,197,122 images
● 154 GB compressed

26. Background
● Started in 2010
● 14,197,122 images
● 154 GB compressed
Tasks
● Object localization for 1000 categories.
● Object detection for 200 fully labeled categories.
● Object detection from video for 30 fully labeled categories.

27. 2012: Alexnet started the deep learning craze.

28. 2014: Google introduced the Inception architecture
Convolution
Pooling
Softmax
Other

29. 2015: Microsoft research introduced Deep Residual
Networks (ResNet)

30. 2017: Momenta introduced Squeeze-and-Excitation
Networks (SENet)

32. kaggle
http://www.kaggle.com

33. ● Created in 2010
● Acquired by Google in 2017
● Competitions support advances in machine learning
and AI in both industry and research
● Clients include

34. kaggle Walkthrough

43. https://www.kaggle.com/learn/overview

44. Q&A

45. Titanic: Machine Learning from Disaster
● Beginner friendly binary classification problem
● Predict whether a person will survive
● Excellent tutorials in python and R
● Join the Local Team!

46. Titanic Competition