3 CP, SoSe 2016

Carlos Garcia C.

garcia@tk.tu-darmstadt.de

based on slides by

Dr. Leonardo Martucci and Florian Volk

Telekooperation Seminar

What? Read and analyze current scientific

publications

Topics: Deep Learning and SDNs

Carlos Garcia C., Telekooperation 2

General Information

How? Select a topic and study it

Write a short report

Review other reports

Present your report

Who? BSc, MSc and Diploma students from Computer Science

Electrical Engineering

and related areas

Carlos Garcia C., Telekooperation 3

General Information

Why? Introduction to a research area Learn to read and analyze

scientific material Present your evaluation

Language? English Even though your advisor might speak

German, your report has to be in English

When? April 11 (now) Introduction Topic presentation Tutorial: Working with Literature

April 19 Topic selection deadline

May 24 First version of your paper

May 31 Submission of review

July 5 First version of your presentation Final version of your survey

July 12 Presentation of your work

Meetings with your advisor

1. Pick a topic, read the provided literature

and find more literature

2. Write an overview or state-of-the-art survey

3. Peer-Review process

Your report will be reviewed by a colleague and by your advisor

You will review a colleague‘s report

4. Correct and improve your report following the

reviewer‘s comments

5. Give a presentation on your report

Carlos Garcia C., Telekooperation 4

5 Steps to Success

Read

Literature

Write Report

Peer review

Correct Report

Presentation

enough

Yes

No

You get 3 graded credit points for Your report: 10 - 12 pages IEEE transactions style paper

(find templates on the course web page)

Your peer review

Your presentation: 12 minutes + 5 minutes of discussion

Carlos Garcia C., Telekooperation 5

Evaluation and Grading

You need to pass all parts!

60 %

Report

15 %

Review

25 %

Presentation

Seminar Topics

Dark SDN (Anonymous SDN)

SDN as the problem Shift of network applications to SDN More private data in SDNs Today’s anonymization services do not fit

SDN as the solution Easy to control & re-route messages Even decoupling from IP address Example:

Black SDN for the Internet of Things, Chakrabarty, 2015

Task(s) Introduce SDNs with anonymity in mind Survey papers related to SDN + anonymity &

privacy Elaborate on how the SDN could protect

anonymity

Jörg Daubert, Telekooperation Lab 7

Motivation (I) Software-Defined Networking (SDN) decouples the control plane from the

data plane and aims at providing a more flexible and cost-effective way to manage tomorrow’s networks

Julien Gedeon, Telekooperation Lab 8

Northbound APIs for SDN

Source: Kirkpatrick (2013)

Motivation (II) While OpenFlow exists as a quasi-standard exsists for the Southbound API

work is still in progress for Northbound APIs

Objectives Gain a deep understanding of SDN and the challenges involved

Task Get a general understanding of what SDN is about

Define what a Northbound API should provide

Get an overview of available SDN controllers

Review what Northbound APIs these controllers use and what capabilities they have

Your own ideas for an API? / Outlook on future work…

Julien Gedeon, Telekooperation Lab 9

Northbound APIs for SDN

Motivation Classic machine translation systems rely on big translation tables

New approaches use thought vectors (more like humans use language) I.e., if you want to translate something into another language you should:

1. understand what the sentence in the first language means

2. and then say the same thing in the other language

instead of translating single words or short sequences

Objectives An understanding of how the current iteration of neural networks accomplish

machine translation Including subparts like word embedding (Word2Vec)

Task A survey about the state-of-the-art approaches for machine translation using

deep learning (thought vector approach)

Timo Nolle, Telekooperation Lab 10

Machine Translation Using Thought Vectors

Motivation Speech recognition is gaining a lot of traction (Siri, Google Now, …)

Current approaches are heavily based on neural networks

Recurrent neural networks (RNNs) can model temporal aspects of speech Dominant approach: Hidden Markov Models (HMMs)

Objectives Get familiar with: Speech recognition / acoustic modelling

Deep learning (RNNs, LSTMs)

Task Find and survey literature on DNN-HMM (and variants) and contrast with

end-to-end approaches using RNNs/LSTMs

Hands-on: Try out EESEN (github.com/srvk/eesen) and/or KALDI (kaldi-asr.org) Both have learning recipes which you can just run.sh without much hassle

Benjamin Milde, Telekooperation Lab 11

Recurrent neural networks (RNNs) in speech recognition

Motivation

Objectives Understand how Deep Learning is used to play top level Go

Explain the Deep Learning techniques used

Task A tutorial paper showing how Deep Learning is used to develop Go bots

Carlos Garcia, Telekooperation Lab 12

Deep Learning can play Go / Baduk / Weiqi

Go is a simple game, yet it is deep and (mathematically) complex

It was thought that computers would never be able to play high level Go

Motivation The ability to create (art) is only associated to humans

Can computers create art?

Objectives Understand how Deep Learning is used in the creation of artistic content

Explain how Deep Learning splits style from content

Task A tutorial paper explaining Deep Learning as applied to artistic style imitation

Carlos Garcia, Telekooperation Lab 13

Deep Learning for the Creation of Artistic Content

Carlos Garcia C., Telekooperation 14

Overview on Topics

ID Supervisor Topic

1 Jörg Daubert Dark SDN

2 Julien Gedeon Northbound APIs for SDN

3 Timo Nolle Machine Translation Using Thought Vectors

4 Carlos Garcia Deep Learning can play Go

5 Carlos Garcia Deep Learning for the Creation of Artistic Content

6 Benjamin Milde Recurrent neural networks (RNNs)

in speech recognition

To select a topic fill out the following form https://form.jotformeu.com/61012268451346

by

Leonardo A. Martucci

Sascha Hauke

Florian Volk and

Carlos Garcia C.

How to work with Literature and write Scientific Material

CONTENT

What’s a scientific publication?

Finding (good) references

Correct referencing

Writing your own paper

Peer-Reviews

* parts of this slide set are based on material provided by Guido Rößling

Scientific Publication a message With scientific background

Offer a new insight of a scientific problem Solution

Problem

Criticism

OR a survey of a research field

Scientific Survey Quality of the survey directly proportional

to the quality of the surveyed material

Doing a survey is about balancing exploitation and exploration Exploit: Continue digging into good literature

Explore: Try to find new things

Carlos Garcia C., Telekooperation

What’s a scientific publication?

01

Books Survey (mostly) about a topic

Journal Articles

Collection of related topics into one magazine (the journal) Quality mostly depends on the Journal Rankings: http://www.core.edu.au/index.php/ Good Journal Good Article

Conferences and Symposia The most recent research achievements Strict page limits Papers followed by a presentation Quality is usually connected to the Conference Rankings: http://www.core.edu.au/index.php/ Good Conference Good Paper

Workshops Mostly for work in progress Good for discussing new ideas

Carlos Garcia C., Telekooperation

Types of Publications

02

Refer back to the original source of information For others to identify the foundations of your work

Giving credit, when credit is due

Not doing so is REALLY bad practice A.K.A. plagiarism

Grundregeln der wissenschaftlichen Ethik am Fachbereich Informatik

Carlos Garcia C., Telekooperation

References and Referencing

03

Scientific publications Articles, papers, books

Standards RFC, ITU, IEEE, W3C etc.

+ All other non-scientific sources Surveys

Magazines

Reports

Can I reference Wikipedia?

NO: no reliable (or stable) information sources

Carlos Garcia C., Telekooperation

What should I reference?

04

1. First, define the message Objective of your publication

define the area of research

2. Read the related work Define the work around your work

Finding out what has been done

3. Survey the related work Evaluate differences Identify trade-offs Create a Storyline

4. Write your publication

Carlos Garcia C., Telekooperation

Writing a Scientific Publication

05

Storyline What is the plot of your survey?

Connect each idea with the next in a sensible way

Find common links

Finding the message The most difficult part (!)

Also, the creative one

go beyond the state of the art

Communicate your message with science

Find the scientific foundations

Identify the challenges

Carlos Garcia C., Telekooperation

1. Your Work, Your Message

!

06

Related Work? Where? For the initial literature ask your supervisor

it will give you a broad idea about the area

Check publication repositories

ACM Digital Lib http://portal.acm.org/portal.cfm

IEEE Xplore http://ieee.org/portal/site

Google Scholar http://scholar.google.com

Academic Search http://academic.research.microsoft.com/

Conference directories http://www.dblp.org/search/

Authors’ home pages

Other sources from the reference lists

REPEAT

Carlos Garcia C., Telekooperation

2a. Related Work? Where? How?

07

Related Work ∞ Identify the relevant sources

Evaluating the importance of a publication

1. Read the abstract

2. Check the reference list

3. Read the conclusions

4. Read the rest

Related work will Compare your results against their results

Be used as input for a survey

Carlos Garcia C., Telekooperation

2b. Related Work and Relevance

Good

Good

Good

Paper Read

Next Paper

No

Yes

Yes

Yes

08

A reference looks like this:

there are also other reference styles

if you use LaTeX to write your report, have a look at BibTeX.

Carlos Garcia C., Telekooperation

Referencing: doing it right

authors

title

how was it published (proceedings) publisher date page number

09

Always have a good paper structure Organize your ideas

Organize your papers

Define it BEFORE starting to add text

Plan the content of each section

Writing skills No one learns without doing it

General Guidelines:

Be concise

Be precise

Carlos Garcia C., Telekooperation

4. Write your Publication

10

Carlos Garcia C., Telekooperation

Peer-Reviews

11

What to Write 1. Short summary of the survey

2. Strong aspects of the work What did you like?

Anything exceptional worth mentioning?

3. Weak aspects of the work Anything wrong that caught your attention immediately?

4. Recommended changes to be made Offer suggestions to improve what you have read

It is NOT the reviewer’s work to … correct the publication

… point out all typo-s, grammar and spelling mistakes

All bullets necessary for full points

A scientific publication is a message; a validated claim

Refer to the original source of information, avoid plagiarism

The peer-review should help, not criticize

Carlos Garcia C., Telekooperation 12

Summary

Good

Good

Good

Paper Read

Next Paper

No

Yes

Yes

Yes

1. Read the abstract

2. Check the reference list

3. Read the conclusions

4. Read the rest