You are on page 1of 21

DEEP TECH TALK :: 2020.07.

25

MENGENAL
> MACHINE 01

> LEARNING

ARRA'DI NUR RIZAL


ABOUT ME
ARRA'DI NUR RIZAL

> SYSTEM DEVELOPER


> SOLUTION & SYSTEM ARCHITECT
> COMPUTATIONAL LINGUIST 02
> SEARCH ARCHITECT

AFFILIATION

> LUND UNIVERSITY


DEEP TECH

> UPPSALA UNIVERSITY


> POLYGLOT INDONESIA
TALK

> ZYWIE LABORATORIUM


> INACL
KNOW THE
DIFFERENCES

03
DEEP TECH
TALK

source:
https://serokell.io/blog/ai-ml-dl-difference
https://www.ospreydata.com/2020/07/06/getting-to-know-ai-and-ml-models/
https://medium.com/ai-in-plain-english/data-science-vs-artificial-intelligence-vs-machine-learning-vs-deep-learning-50d3718d51e5
AI VALUE CREATION BY 2030

MOST AFFECTED
$ 13 TRILLION (Mckinsey, September 2018)

Job with repetitive task.


i.e retail worker

LEAST AFFECTED

04 Highly manual job. i.e


hairdresser or surgeon

FOR COMPANIES

Skilling and reskilling source: https://www.mckinsey.com/featured-insights/artificial-intelligence/notes-from-the-ai-frontier-modeling-the-impact-of-ai-on-the-world-economy#

people to work with AI


DEEP TECH TALK
AI TYPES AGI does not exist yet. Most
"AI" is Supervised Machine
Learning. Learns to predict
output B from an input A

05

DEEP TECH TALK


Source: Andrew Ng’s new course: “AI for everyone” (week 1)
ML APPLICATIONS
HEALTHCARE
Design treatment plans, Predict illness. etc
State-of-the-art: Artificial Intelligence Outperforms
Doctors in Breast Cancer Diagnosis

RETAIL AND COMMERCE


06
Suggest product, Predict customer behaviour
Usages are typically on the procurement, sales and
marketing

TRANSPORTATION
Mobility and Driver assistant
The most famous usages is in the self-driving car, but
it also uses in cruise control, self-parking, and
traffic management.
Source: Nature, Ericsson
AMOUNT OF DATA WHY NOW ?
We have a lot more data
to be used for training

COMPUTING POWER
07 Model training is
accelerated by GPU usages

NEW EXCITMENT

Breaktrough brings
excitment in the AI
research domain
DEEP TECH TALK
Source: Andrew Ng’s new course: “AI for everyone” (week 1)
DEMISTIFYING AI

08

CAN AND CAN'T

Source: Andrew Ng’s new course: “AI for everyone” (week 1)


09

MACHINE LEARNING VS Data Science can visualize the


DATA SCIENCE past and Machine Learning can
predict the future.
Source: Andrew Ng’s new course: “AI for everyone” (week 1)
10

Example: Supervised Deep


MACHINE LEARNING IN A
Learning
NUTSHELL

Source: Andrew Ng’s new course: “AI for everyone” (week 1)


DEEP TECH TALK
MACHINE LEARNING
TERMS

TRAINING
Process where the program is trying to fit the data to the
target
11
MODEL
The weights we get after training

INFERENCE
Using trained model to predict an output

OVERFIT
The model is not generalized enough
DEEP TECH TALK
ML TYPES
> LEARNING PROBLEMS

SUPERVISED LEARNING
Map input to target variable. Algo: Nearest Neighbour,
Naive Bayes, Decision Trees, Linear Regression, Support
12
Vector Machine (SVM), Neural Networks

UNSUPERVISED LEARNING
Extract relationship in data. Algo: K-means, Word2Vec

REINFORCEMENT LEARNING
Agent operates in an environment and must learn to operate
using feedback. Algo: Q-Learning, Temporal Difference
(TD),Deep Adversarial Networks
DEEP TECH TALK
ML TYPES
> HYBRID LEARNING PROBLEMS

SEMI-SUPERVISED LEARNING
Supervised + Unsupervised :: small labeled data and use
13 the rest unlabeled data in unsupervised manner.

SELF-SUPERVISED LEARNING
Unsupervised learning framed as supervised learning. so we
can use supervised algo. i.e. colorization task : make
images grayscale to predict the color.

MULTI-INSTANCE LEARNING
Supervised learning where data-point is group and the
group is labeled instead of the data-point.
DEEP TECH TALK
ML TYPES
> LEARNING TECHNIQUES

MULTI-TASK LEARNING
Supervised learning which address multiple related problem
using the same dataset.
14
ACTIVE LEARNING

Technique where model can ask human during training


process

ONLINE LEARNING
Update model before prediction or after last observation
DEEP TECH TALK
ML TYPES
> LEARNING TECHNIQUES

TRANSFER LEARNING

Trained model in one task and use it as starting point in


15 other task

ENSEMBLE LEARNING

Combine results from multiple models to do prediction


ACQUIRING DATA

16

STRUCTURED DATA
&
UNSTRUCTURED DATA

garbage in - garbage out


Source: Andrew Ng’s new course: “AI for everyone” (week 1)
STRENGTHS WEAKNESSSES

ML tends to work well when: ML tends to work poorly


1. Learning a "simple" when:
concept. When human took 1. Learning complex concepts
17 less than 1 second to do from small amounts of
2. Has a lot of data data
2. It is asked to perform on
new types of data

Source: Andrew Ng’s new course: “AI for everyone” (week 1)


WHERE TO
LEARN

ONLINE COURSES 18
easy to start, sufficient for motivated
student

BOOTCAMP
learns practical skill

UNIVERSITIES
deep dive into the core DEEP TECH TALK
PRIORS
LINEAR ALGEBRA

CALCULUS

PROBABILITY THEORY
19 OPTIMIZATION THEORY

ALGORITHMS

PROGRAMMING LANGUAGE

ENGLISH
LINGUISTICS (NLP)

DEEP TECH TALK


Source: https://www.researchgate.net/figure/A-3D-surface-plot-demonstrating-local-and-global-minima_fig10_305881922
TOOLS

PROGRAMMING LANGUAGE: PYHTON

LIBRARIES: SCIKIT-LEARN, PANDAS


20
FRAMEWORKS: TENSORFLOW, PYTORCH, KERAS, THEANO,
ACCORD.NET, APPLE'S CORE ML, GOOGLE ML KIT,
TENSORFLOW.JS

APLICATIONS: KNIME, WEKA, RAPIDMINER, APACHE SPARK MLLIB

EDITOR: JUPYTER NOTEBOOK, GOOGLE COLAB

DEEP TECH TALK


DEEP TECH TALK

QUESTIONS ?

anrizal@gmail.com

You might also like