Generative Adversarial Nets
Seminar Report
Submitted in partial fulfillment of the requirements for the award of B.Tech Degree in Computer Science and Engineering of A P J Abdul Kalam Technological University
Submitted by
Name : Santhisenan A Roll no : 50 Semester : S7 Student ID: TVE15CS050
Department of Computer Science and Engineering C OLLEGE OF E NGINEERING T RIVANDRUM
Student ID: TVE15CS050
Dept Of CSE
Declaration I undersigned hereby declare that the seminar report Generative Adversarial Networks, Networks, submitted for partial fulfillment of the requirements for the award of degree of Bachelor of Techn echnol olog ogy y of the the APJ APJ Abdu Abdull Kala Kalam m Techn echnol olog ogic ical al Unive nivers rsit ity y, Kera Kerala la is a bona bona fide fide work work done done
by me under under superv supervis isio ion n of Dr This subm submis issi sion on repr repres esen ents ts my idea ideass in Dr.. Ajeesh Ramanujan Ramanujan. This my own words and where ideas or words of others have been included, I have adequately and and accu accura rate tely ly cite cited d and and refe refere renc nced ed the the origi origina nall sour source ces. s. I also also decl declar are e that that I have have adhe adhere red d to ethi ethics cs of acad academ emic ic hone honest sty y and and inte integr grit ity y and and have have no nott misr misrep epre rese sent nted ed or fabri fabrica cate ted d any any data data or idea or fact or source in my submission. submission. I understand that any violation of the above will be a cause for disciplinary action by the institute and/or the University and can also evoke pena penall acti action on from from the the sour source cess whic which h have have thus thus no nott been been prop proper erly ly cite cited d or from from whom whom prop proper er
permi permissi ssion on has no nott been been obtai obtaine ned. d. This This repo report rt has not been been prev previou iously sly forme formed d the basis basis for for the award of any degree, diploma or similar title of any other University.
Santhi Santhisen senan an A Department of Computer Science and Engineering College of Engineering Trivandrum
i
Student ID: TVE15CS050
Dept Of CSE
Departm Dep artment ent of Compute Computerr Science Science and Enginee Engineering ring C OLLEGE OF E NGINEERING T RIVANDRUM
Certificate This is to certify that the seminar titled "Generative Adversarial Networks" submitted Networks" submitted by Santhisenan A (TVE15CS050), (TVE15CS050) , to College of engineering Trivandrum, for the award of the degree of Bachelor Bachelor of Technology in Computer Science and Engineering, of A P J Abdul Kalam Technological University during the year 2015-2019., 2015-2019., is a bona fide record of the work done by him under our supervision. The contents of this report, in full or in parts, have not been submitted to any other Institute or University for the award of any degree or diploma.
Dr. Ajeesh Ramanujan Dr. Ramanujan Assistant Professor Dept. of Computer Science College of Engineering Trivandrum
ii
Student ID: TVE15CS050
Dept Of CSE
Acknowledgeme Acknowledgement nt Firs Firstl tly y, I woul would d like like to than thank k the the almi almigh ghty ty for for givi giving ng me the the wisd wisdom om and and grac grace e for for pres presen enti ting ng my seminar. With a profound sense of gratitude, I would like to express my heartfelt thanks to my guide Dr. guide Dr. Ajeesh Ramanujan, Ramanujan, Assistant Professor, Department of Computer Science and Engine Engineeri ering ng for his expert expert guida guidance nce,, co-ope co-operat ration ion and immens immense e encour encourage ageme ment nt in pursui pursuing ng this seminar s eminar.. I am very much thankful to Dr to Dr.. Shreeleksh Shreelekshmi mi R, R, Head of the Department and Prof. and Prof. Vipin Vasu A V , Assi Assist stan antt Prof Profes esso sorr of Depa Depart rtme ment nt of Comp Comput uter er Scie Scienc nce e and and Engi Engine neer erin ing, g, for for providing necessary facilities and his sincere co-operation. My sinc sincer ere e thank thankss is exte extend nded ed to all all the the teac teache hers rs of the the depar departm tmen entt of Comp Comput uter er Scie Scienc nce e and to all my friends for their help and support.
Santhise Santhisenan nan A
iii
Student ID: TVE15CS050
Dept Of CSE
Abstract Abstract Generative adversarial nets or GANs can be thought of as a game theoretic approach to deep learning. Every GAN will have two major components, a generator G, and a discriminator D. They are like two adversaries, playing against each other in a game. G tries to generate images that look real from noise, and D tries to identify fake images generated by G. G and D are in a constant battle throughout the training process. G tries to fool D by making real ealisti isticc fake fake imag images es and and D trie triess no nott to get get fool fooled ed by G. It is impo import rtan antt to have ave good generator, otherwise generator can never fool the discriminator and mode modell may may no nott co conv nver erge ge.. It is equa equall lly y impo import rtan antt to have have a good good disc discri rimi mina nato torr, as otherwi rwise images ges gener nerated by the GAN wil will be of no use. One of the most ost impo import rtan antt cha challen llenge gess to GAN GAN is that that if any any on one e syst system em fails ails,, the the whol whole e syst system em fails. A variety of GANs like WGAN, CGAN, InfoGAN, SRGAN etc. has been develope developed d for various various applications applications.. GANs is an area area in which which very active active research research is being done. Keywords: Generative algorithms, Discriminator, Discriminator, Generator
iv
Contents Declaration . . . . Certificate . . . . . Acknowledgement Acknowledgement Abstract . . . . . . . Abbreviations Abbreviations . . . Notations . . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. i . ii . iii . . iv . vi . vii
1 Generative Adversarial Nets 1.1 Generative Adversarial Networks . . . . . . . . . . . . . . . . . . . 1.2 1.2 Gene Generrativ tive vs Discr iscrim imin ina ative ive algo algori ritthms hms . . . . . . . . . . . . . . 1.3 How GANs work . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
v
1 1 1 1
Student ID: TVE15CS050
Dept Of CSE
Abbreviations Abbreviations • GAN Generative Adversarial Adversarial Network
vi
Student ID: TVE15CS050
Dept Of CSE
Notation
vii
Chapter ter 1 Genera Generativ tive e Adve Adversa rsaria riall Nets 1.1 Gener Generat ativ ive e Adve Adversa rsaria riall Networ etworks ks Generative adversarial networks (GANs) are a class of artificial intelligence algorithms used in unsupervised machine learning, implemented by a system of two neural networks networks contesting with each other in a zero-sum game framework. They were introduced by Ian Goodfellow et al. in 2014[ 1]. This tech techni niqu que e can can gener generat ate e phot photog ogra raph phss that that look look at leas leastt supe superfic rficia iall lly y auth authen enti ticc to human observers, having many realistic characteristics.[2] characteristics.[2]
1.2 1.2 Gene Genera rati tive ve vs Di Disc scri rimi mina nati tive ve algo algorit rithm hmss To understand GANs, you should know how generative algorithms work, and for that, contrasting contrasting them with discriminative discriminative algorithms is instructive. Disc Di scri rimi mina nati tive ve algo algori rith thms ms try to clas classi sify fy inpu inputt data data;; that that is, is, give given n the the feat featur ures es of a data data insta nstanc nce e, the they pred predic ictt a label abel or cate catego gory ry to whi which that that data data bel belon ongs gs.. Discriminative Discriminative algorithms map features features to labels. They are concerned concerned solely with that correlation. One way to think about generativ generative e algorithms is that that they they do the the oppo opposi site te.. Inste nstead ad of pred predic icti ting ng a labe labell give given n cert certai ain n feat featur ures es,, they attempt to predict features given a certain label.
1.3 How GANs work One One neur neural al netw networ ork, k, call called ed the the gene genera rato torr, gene genera rate tess new new data data inst instan ance ces, s, whil while e the the othe otherr, the the disc discrim rimin inat ator or,, eval evalua uate tess them them for for auth authen enti tici city; ty; i.e. i.e. the the disc discrim rimii1
Student ID: TVE15CS050
Dept Of CSE
nator decides whether each instance of data it reviews belongs to the actual training dataset or not.[3] Steps taken by a GAN are:
• The generator takes some random noise and returns an image • This generated image is fed to the discriminator along with a stream of images taken the actual dataset • The discriminator taken in both real and fake images and returns the probability of the fake image being real.
2
Bibliography [1] Goodfellow Goodfellow, Ian; Pouget-Abadie, Pouget-Abadie, Jean; Mirza, Mehdi; Xu, Bing; WardeFarley arley,, David; David; Ozair Ozair,, Sherj Sherjil; il; Courvil Courville le,, Aaron; Aaron; Bengio Bengio,, Joshu Joshua a Generative Adversarial Networks Networks 2014 [2] Salim Salimans ans,, Tim; Tim; Goodfel Goodfello low w, Ian; Ian; Zaremb Zaremba, a, Wojciec ojciech; h; Cheung, Cheung, Vicki; Vicki; RadRadford, Alec; Chen, Xi. Improved Xi. Improved Techniques for Training GANs 2016 [3] https://deeplearning4j.org/generative-adv https://deeplearning4j.org/generative-adversarial-network ersarial-network
3
