Machine Learning Wars:
Deep Learning vs GBM
Sefik Ilkin Serengil
software developer @ softtech
sefiks.com
serengil
Sefik Ilkin Serengil
Software Developer @ ‘10
CS graduate, MSc @ ‘11
Instructor @
Gradient
Boosting
Deep
Learning
Deep Learning
gives us
superpowers
We can recognize face identities with Deep Learning
https://youtu.be/tSU_lNi0gQQ
Or we can find facial expressions and emotion
https://youtu.be/Y7DfLvLKScs
We also know that what would it look
like if Vincent van Gogh had painted
Galatasaray University
https://sefiks.com/2018/07/20/artistic-style-transfer-with-deep-learning/
https://sefiks.com/2017/12/10/transfer-learning-in-keras-using-inception-v3/
We have outcomes of Tech Giants in seconds through transfer learning
It works everytime?
Microsoft developed its bot Tay.ai with deep learning
It becomes a racist one in hours
This is because we cannot respond why and
how questions in deep learning as well…
Decisions in deep learning are basically made by matrix multiplication
and this is not understandable by human
We might move the hammel with an alternative algorithm
If the decision is wearing rain jacket, we can read reasons backwardly
https://archive.ics.uci.edu/ml/datasets/car+evaluation
We check just a single feature in same level of tree
Decision tree consists of 700
lines of if statetements!
Luckily, decision tree algorithm
can exctract these rules and
developers wouldn’t write
these if statements.
Decision Trees are
non-linear ML models
Classify this data set
Linear Model
E.g. Perceptron, Linear SVM
Decision Trees
fails! They can handle
Now, classify this data set
Linear Model Decision Tree
Simple linear model can easily classify this task Decision trees still can model the problem, but it
wouldn’t be easy!
A single tree is not
strong enough
Gradient Boosting Machines (GBM)
Gradient Descent + Boosting
Gradient Boosting Decision Trees (GBDT)
Multiple Additive Regression Trees (MART)
Gradient Boosting Regression Trees (GBRT)
How gradient
boosting works
100 m
70 m
30 m
15 m
15 m
20 m
-5 m
5 m
https://playground.tensorflow.org
http://arogozhnikov.github.io/2016/07/05/gradient_boosting_playground.html
Weak classifiers can come together and create a
strong classifier
XGBoost: A Scalable Tree Boosting System
https://www.kdd.org/kdd2016/papers/files/rfp0697-chenAemb.pdf
2015 KAGGLE GRAND PRIX AMONG 29 WINNING SOLUTIONS
Podium Ceremony
1
2 3
GBM
17 solutions
Deep Neural Nets
11 solutions 9 solutions
GBM + Neural Nets
GBM Frameworks
Accuracy
https://youtu.be/8o0e-r0B5xQ?t=266
Adoption
https://www.quora.com/Does-the-new-CatBoost-algorithm-by-Yandex-outperform-the-infamous-XGBoost-in-Kaggle-competitions
LightGBM is 10 times faster than XGBoost!
https://github.com/szilard/GBM-perf
LightGBM performs faster in CPU whereas XGBoost
runs faster in GPU. Consider the hardware you have
before adopting a framework
What data science methods are used at finance industry?
https://www.kaggle.com/surveys/2017
Decision trees are more
popular in data science world
My favorite machine learning is Neural Nets. That’s
my favorite.
My 2nd favorite machine learning is SVD.
Everyone says, oh don’t you prefer gradient boosted
trees? I know GBTs are great but I like NN best and I
like SVD next best.
Francois Chollet
Creator of Keras, Google AI Researcher
https://twitter.com/fchollet/status/992465859904876544
Thank you for your attention!
sefiks.com serengil

Machine Learning Wars: Deep Learning vs Gradient Boosting Machines