Machine Learning

CS-433

Martin Jaggi & Nicolas Flammarion

21st Sept 2021




 Alternatives
Master
EE-559 – Deep Learning Bachelor
CIVIL-459 – Deep learn. for autonomous vehicles CS-233a – Intro to ML
EE-411 – Fundamentals of inference and learning CS-233b – Intro to ML
MGT-418 – Convex optimization CS-330 – Artificial Intelligence
MATH-403 – Low-rank approximation techniques BIO-322 – Intro to ML for bioengineers
MATH-412 – Statistical ML
MICRO-455 – Applied ML
MICRO-401 – ML Programming Seminars, Doctoral Courses and continued education
DH-406 – ML for Digital Humanities
ENG-704 – EECS Seminar: Advanced Topics in ML
CS-430 – Intelligent Agents
CS-723 – Topics in ML Systems
CS-439 – Optimization for ML
EE-608 – Deep Learn. for Natural Language Proc.
CS-401 – Applied Data Analysis
EE-618 – Theory and M. for Reinforcement Learning
CS-449 – Systems for data science
EE-613 – ML for engineers
EE-556 – Mathematics of data
EPFL Extension School – Applied Data Science: ML
CS-526 – Learning theory
MICRO-570 – Advanced ML
CS-433 – Machine Learning
Talks: join mailing list:
[email protected] .ch

fl

 Course Logistics
Assessment

✤ Project 1 (10%), due Nov 1st

✤ Project 2 (30%), due Dec 23th
✤ Final exam (60%)

Respect @ EPFL
 5

Respect @ EPFL

Safe Space https://go.epfl.ch/safespace

 Course Logistics MJ

Lectures NF

tentativ
schedule

MJ
e

 Course Logistics
Lectures

Tuesday 2x45mins, Room: Rolex learning center

Thursday 2x45mins, Room: SG1

We provide PDF lecture notes on our webpage and GitHub, and

streaming&recordings of all lectures on YouTub

 Course Logistics
Exercises

Thursday 14:15 - 16:00 - live interaction!

Room alternatives: real or zoom
sign up on sheet to see assignment

All labs and projects are in Python.

See the rst lab to get started.

Code Repository for Labs: github.com/epfml/ML_course

fi









 Course Logistics
Team of assistants contact us: online forum!

Maksym Andriushchenk Raphaël Attia

El Mahdi Chayt
Axel Dinh Van Ch
Lie H
Anastasiia Koloskov Karim Hadidan
Tao Li Xiaowen Jian
Amirkeivan Mohtasham Ella Rajaonso
Ehsan Pajouheshga Ekrem Yilmaze
Scott Pesm
Yingxue Y
Thijs Vogel
Tianzong Zhang Oguz Yuskel
e

 Course Logistics
Projects

✤ Project 1 (10%), due Nov 1st

✤ Project 2 (30%), due Dec 23th

Real-world problems, Python, Groups of 3 Students

 What to expect?

✤ overview over M
✤ basic understanding of most important ML methods
and fundamental concept
✤ experience how ML is done on a practical problem
L

 What not to expect?

✤ instantly becoming an expert data scientist / ML

researcher / statistician / large-scale ML specialist /
ML software expert / domain exper

✤ learn about that newest hack you saw on blog xyz

✤ build your own self-aware AI

t



Your colleagues here
EPFL sections
 Introduction

What is Machine Learning?

What is Machine Learning?

algorithms that can

learn from data

 traditions…

data
traditional
output
programming
rules

data
ML rules
desired
output ML mode
weights w
l

 traditions…

data M

data M

ML mode
weights w




l

 Learning Functions from Data

pixels
input output
image
hussar monkey
classi cation

“Bonjour! Comment
text

“Hello! how are you?” translation

allez-vous?”
audio

“Hello! how are you?” speech to text

pixels

“a dog is sitting at the beach image

next to another dog” captioning

writing
text

wha…t?
assistant
fi




 input output
“moon landing webpage 1
webpage web search
conspiracy”
webpage 3
browsing history on recommender
fashion website system

medical image
pixels

melanoma
processing

“look at whether it works for

video

lip reading
the UK or not”
The bot must be trained in a language capable of decoding Python's
strings and displaying it on a high quality display, in order to be able
“moon landing to produce what they have learned in English, and indeed, these
images, have been uploaded to the web for quite some time. If this
type of thing is indeed present in the wild, then what sort of wild bot
should I be worried about text generation
conspiracy…” Thanks to this one speci c experiment performed on the same day - as
described by the author
Using some kind of neural network to learn speech, and being able to
decode it in order to communicate with others (including yourself)
through its GU

professional dance
dance transfer
+ photo of myself https://youtu.be/PCBTZh41Ris?t=139
I

fi


?



image source
 Classi cation xi 2 R d

Training data
fi
 The Learning Algorithm xi 2 R d

(Stochastic
Gradient
Descent)
x
w := w + ·x
iteration cost: O(d)

Perceptron Support-Vector-Machine
(Rosenblatt 1957) (Cortes & Vapnik 1995)




 towards…
understanding intelligence
?

if-then-else
≠
intelligence







 towards…
understanding intelligence
?
Machine Learnin Neuroscience / HB

vs


g

 What is the difference between

Artificial Intelligence, Data Mining,
Statistics, Machine Learning?

 AI X

Computer Statistics
Science Machine
Learning

Application
domain



 Job Skills

X
Machine Learning

Coding Math &

Skills Statistics
Data
Scientist
Software Traditional
Engineer Scientist

Application
domain



 Cycles of popularity

Machine Learnin
Linear Algebra

source
g

 why ML?

Applications
 Industry Applications

✤ majority of industries, originally not ‘digital’

✤ agriculture, NGOs, ‘sharing economy', logistics,
delivery, services, manufacturing, sports,
personalized health, call centers, entertainment, …

✤ not only the

‘usual suspects


’



 Applications in Other Sciences

✤ increasingly data drive

✤ … Psychology, Economics, Medicine, Social science
✤ science of X ➔ digital science of X

en.wikipedia.org/wiki/Index_of_branches_of_science
n

ML is only a small part!

 History
✤ ML is not new!

London 185
cholera outbreak
4

 History

✤ ML is not new!

✤ the early days - 1950ies and 1960ies

✤ Neural network
✤ Turing
s

 “the emb
electroni r y o of an
What has changed c compu
t e r that …
will be a
ble to wa
see, writ lk, talk,
3
1950s: 10 FLOPS e, reprod
and be co u ce itself
18 nscious o
2021: 10 FLOPS existence f its

.” 1958

 Challenges
✤ Hyp
✤ cycles of AI popularit

✤ Data Ethics, Privacy, Fairness

e

Challenges
 Challenges
✤ Hyp
✤ cycles of AI popularit

✤ Data Ethics, Privacy, Fairnes

✤ Lack of Interpretabilit
✤ example: medical applications of deep learnin

✤ Social Implications of AI,

Threats from Super-human AI
✤ see Nick Bostrom, Yuval Harari

need: Scientific Method, Reproducible

Research, Open Source and Open Data
e



s

 ML Applications
by CS- Master Students
4
3
3


 ML science projects by your colleagues
Machine Learning-based Estimation of Cardiac Contractility from Peripheral Pressure Wavefor Regularized maximum likelihood estimation – TRANSP-O
Deep learning techniques for geometric matching of C. Elegans brain microscopy image Stroke Level Estimation through pac-man game data played by acute stroke patient
Benchmarking Machine Learning Methods for Eukaryote/Prokaryote Contigs Classi catio STLM: Steganography in Text using Language Model
Machine Learning for Science: Classi cation of Skin Samples Using Mass Spectrometry Analysi Eastern Rituals Search Engine (ERSE
Application of Deep Knockoffs for fMRI to Generate Surrogate Dat Cough Classi e
Automatic detection of weak cipher usage in aircraft communication Extracting high value lung ultrasound images from video for the diagnosis and prognosis of COVID-1
Predicting Topic Change and Emoji Usage from Twitter Dat Detecting rooftop available surface for installing PV modules in aerial images using Deep Learnin
Cell Nuclei Segmentation in 2D Fluorescence Microscopy Image Dimensionality reduction and clustering of energy consumption time series in supermarket building
Unsupervised cell classi cation in ow cytometry dat Protein-Protein Interaction
Predicting chemicals concentration in water streams using Gradient Boosting Regresso Predicting gene-gene relationship with CNNC mode
Extracting Masonry Building Facades through Polygon Image Segmentatio PneumoNet: Neural networks for the detection of pneumonia from digital lung auscultation audi
Sequence-dependent clustering of DNA in Protein-DNA Xray crystal data and in cgDNA+ mode Predicting errors during Pacman for stroke patient
Applying the VoxelMorph Framework to C.Elegans Brain Dat Galaxy Detection Machine Learning Projec
Using forearm sEMG to control individual ngers of a robotic han Automatic detection of available area for rooftop solar panel installation
Music beyond Major and Mino Prediction of myocardial infection risk after stenosis diagnosi
Avalanche Forecasting: An Ordinal Regression Approac LC3 compressive strength analysi
Machine Learning for Side-Channel Disassembl Adapting Attention Guided Camera Localization for the Geodetic Engineering Laborator
Multi-object Detection and Trackin Machine learning models to predict the diagnosis and risk of COVID-19 from clinical data in Switzerlan
Motion-based Similarity Search in Videos of Confucian Ritual Facades and Openings Detection Based on Different Deep Learning Model
Detecting the Degree of Cavitation In Situ in Young Tree Variational Inference compared to Markov Chain Monte Carlo for modelling gene expressio
Machine Learning in Chemistr 3D Spatiotemporal clustering of mixed-type medical data in Tanzani
Personalized Federated Image classi cation using Weight Erosio Classi cation and Clustering on Schizophrenic Patient’s Dat
In-crystal Gamma-interaction localization for positron emission tomography (PET) from Cherenkov photon TRANSP-OR – Prediction of mode of transportatio
Classi cation of zebra sh embryo using various ML method Learned cross-domain descriptors (LCD) for drone navigatio
Resource-Ef cient Machine Learning Algorithm Design for On-Implant Neurological Symptom Detectio What if Interactive GlobalCOVID Policy Simulato
Ebola Virus Disease Diagnosis for West African Ebola Virus epidemi Image Segmentation of Adenovirus Particles in Food Vacuoles of Eukaryotic Organism
Supervised classi cation of y behaviors from posetracking dat Music Beyond Major and Mino
Cell-type classi cation from microscope imagin Determining the important features for estimating the reproduction number in the COVID-19 pandemi
COVID-19 Predictions using Machine Learnin Exploring chord embedding spaces between musical composers and era
Unsupervised time series analysis of country wise COVID dat Vector Embeddings of Musical Chord
Voxelmorp Word embeddings and transformer models for optimal learnin
Unsupervised classi cation of video games style Identi cation of re periods from air quality monitoring network measurement
Can the Style and Wording in Critical Reviews of Video Games Predict its PEGI Labelling Drone and pigeon detectio
Ensemble Methods for Dynamic Portfolio Valuatio Characterization of turbulent structures in tokamak
Vector embeddings of harmonies in music with deep learnin Improving Deep Learning models for EMG decoding used for prosthesis control enhancemen
Robustness of U-Net based models to common image artefact Pneumonia Diagnosis based on CNN-LSTM-BERT Mode
Recognizing Humor and Predicting Humor Ratings in Short Text L-form bacteria segmentatio
Segmentation of cell nuclei in 2D microscopy images with CNN Machine Learning for Spaced Repetition in Human Learnin
Mechanism of Action (MoA) Prediction – Kaggle Competitio COVID-19 risk strati cation on Chest X-Rays: performance on a small cohort of patients in Switzerlan
Diagnostic and Prognostic models for Ebol Dry vs Wet Cough Automatic Classi cation using the COUGHVID Datase
Automatic Grading of Handwritten Student Essay Improving Freshwater Quality Measurements through Machine Learnin
Stroke Level Prediction through Pacman Game Dat Lesion detection on cardiology images using Deep Learnin
Among Us Project 2 – Market states prediction 3D to 2D feature matching for next generation 3D mapping algorithm
Calibrate a model of OTC market
fi
fi
fi
h

fi
fi
fi
r

fi
4
fi
fi
fi
fi
fi
s

fl
n

fl
fi
fi
s

fi
)

fi
a

fi
s

2
0
2
0
 Reproducibility Challenge

reproduce (parts of) the experiments in a machine learning research paper

https://paperswithcode.com/rc2021

(the papers can be from any conference such as 


NeurIPS, ICML, ICLR, ACL-IJCNLP, EMNLP, CVPR, ICCV, AAAI and IJCAI

Reproducibility Challenge 2020

On Warm-Starting Neural Network Trainin
Can gradient clipping mitigate label noise
Sanity-Checking Pruning Methods: Random Tickets can win the Jackpo
Learning to Play Sequential Games versus Unknown Opponent
Distributed Distillation for On-Device Learnin
:

 why ML?

Applications, cont.
 Image Data

✤ Astronom
✤ Face recognitio
✤ 2D + 3D medical
imagin
✤ OC
✤ self-driving cars

how-old.net
R

 neutral positive Tom Brady wins AFC offensive player of the week for 22nd time http://t.co/gwjLE1
negative negative Watching Contraband on the PVR & it's too frigging predictable to continue wa

Text Data
negative negative @JoshNorris @Rotoworld_Draft I'd be pretty mad if the Packers took Bernard in th
neutral neutral Herald Sun: AFL stars make their UFC 152 picks: DANE Swan and Gary Ablett give
negative negative Steal by Chalmers, on the break away and is fouled by Garnett. That is his 3rd fou
neutral neutral @Holly_Gilchrist you out again on Thursday for #aNightmareOnGeorgeStreet at Ch
✤ Text Understanding & Text Generatio
neutral positive Free to Watch!!! Justified: Justified follows Marshal Raylan Givens, a modern day 1
neutral neutral @_BigDaddyDouley Come SUPPORT the SHOW/MOVEMENT at Park dale High Scho
✤ Spam Detectio positive positive @drewbrees I admire the relationship U have with your family. Lol iron man's a pr
neutral neutral @shuayb_ well i went maths on mon, tues + wed but cba now youu? &nopee
positive positive Lance just left, dinner with the fam was great. Managed to watch Napoleon Dynam
✤ User Conten positive positive Come see the David Bowie tribute show I'm in @ King King, H'wood, Nov 4 &
positive positive Your like Jordan's on a Saturday I got to have you and I cannot wait. .
✤ Medical Tex negative neutral But i wanna wear my Concords tomorrow though but i don't feel like it
positive neutral Gonna watch Grey's Anatomy all day today and tomorrow(:
✤ Machine Translation negative neutral @CoachVac heey do you know anything about UVA's fallll fest loll they invited me s
neutral neutral @DustyEf when that sun is high in that Texas sky, I'll be buckin it to county fair. A
neutral positive Up 20 points in my money league with Vernon Davis and L. Fitz still to go tomorrow
neutral positive DEEJAYING this FRIDAY in THE FIRST CHOP it's CHRIS actual SMITH with a smash
negative negative The Rick Santorum signing that was scheduled for tomorrow at the Books A Million
positive neutral @dreami9 lol yep looks like it! Was after El Clasico on Sunday. I didn't like her lol a
neutral neutral Back in Stoke on Trent for the 2nd time today!
neutral neutral First Girls Varsity Basketball Game tomorrow at 6:00 pm Then Football Senior nigh
neutral neutral #UFC lightweights @Young__Assassin VS @jamievarner set for TUF 16 Finale on th
neutral neutral @OOOOO_WEEEE slide thru sometime this weekend ill have somethin yu can sip o
negative negative @DannyB618 Sure absolutely-- I meant out of the Bachmann, Perry, Santorum, H
negative negative @RichardGordon48 re Levein discussion on Wed. Can't keep changing boss, but he
neutral neutral Today In History November 02, 1958 Elvis gave a party at his hotel before going o
neutral positive Hustle cause you got to then kick back n party everyday like its Fri
positive positive I can't sleep. Way too exited about Vancouver tomorrow! I'm like a kid at Christma
t

Medical: Genetic Data

ref
 Audio & Multimodal Data

✤ Hearing aid
✤ Voice Recognitio

✤ Automatic Translatio
✤ Lip Readin
✤ Video Analysis
g

 Numerical / Sensor Data

✤ Cer
✤ Astronomy/ Telescopes
✤ Fitness Tracker
✤ Weather Forecas
✤ Robotic
✤ Kinect
n

 Games & Simulations

✤ Immediate Feedbac
✤ Chess, G
✤ Physical World
o

 Internet Data

Products / Words
★★
★
★
★★

Customers / Words
★★

★
>
★★
★★
★
⇡ UV
★★ ★★
★★ ★

★★

★★
★★ ★
★

Recommender systems, Virtual assistants, Ads











New Opportunities?

Your turn
 up next:

✤ Regressio
✤ Linear Regressio
✤ Classi catio
✤ … fundamental concepts of ML
fi
n