Anthill Inside 2017

On theory and concepts in Machine Learning, Deep Learning and Artificial Intelligence. Formerly Deep Learning Conf.


Hitchhiker’s Guide to Generative Adversarial Networks (GANs)

Submitted by Ramanan Balakrishnan (@ramananbalakrishnan) on Saturday, 29 April 2017

Section: Full talk Technical level: Intermediate

View proposal in schedule


Unsupervised learning has always been a tough problem to crack for researchers and practitioners. The last few years, however, have seen huge strides being made here with the widespread development of generative models, or specifically Generative Adversarial Networks (GANs).

GANs are eating the (deep learning) world, would not be an exaggeration when referring to the past year. Research groups and companies have come out with multiple variations, optimization techniques and network architectures to boost the capabilities of GANs.

The aim of this talk is to provide a comprehensive introduction to GANs, understand their working principles, provide an overview of recent developments and also peek into the future that lies ahead.

every project described in the presentation comes with a hyperlinked reference on the bottom-right of the slide - encouraged to click through to know more about each project


Motivation + Overview of Generative Networks [5 mins]

  • Reasoning for interest in generative networks
  • Overview of options for such networks
  • Issues and motivations for developing GANs

GANs - Fundamentals [12 mins]

Introduce the framework under which GANs can be studied. The following ideas will be discussed to enable a clear understanding of the fundaments

  • GANs as games played betweene adversarial networks
  • Generator v. Discriminator networks and “goals” for each network
  • Loss function options and training process for GANs
  • DCGAN architecture

GANs - Applications and Recent Developments [10 mins]

An overview of developments in the past year. Touching upon each of the following, covering how they were developed, salient contributions and applications that can be realized.

  • SRGANs for single image super-resolution
  • Interactive GANs for image generation
  • pix2pix, image to image transalation
  • Conditional GANs for text to image synthesis
  • Image inpainting using DCGANs
  • Using multiple GANs for cross-domain application and “style transfer”

Issues & Improvements [8 mins]

Overview of the issues affecting GANs and the steps that are being taken to combat them.

  • Recent developments that are helping solve some of the known stability problems
  • Tricks to optimize development and training of GANs
  • Looking at impact of GANs going forward and effect on other aspects of generative modelling


Basic understanding of the current capabilities of deep learning

Speaker bio

I am a member of the data science team at Semantics3 - building data-powered software for ecommerce-focused companies. Over the years, I have had the chance to dabble in various fields covering data processing, pipeline setup, database management and data science. When not picking locks, or scuba diving, I usually blog about my technical adventures at our team’s engineering blog.




  • Ananth Krishnamoorthy (@akrishnamoorthy) 2 years ago

    I think this should be a longer talk. This is a good emerging area and will need some time for discussion

    • Ramanan Balakrishnan (@ramananbalakrishnan) Proposer 2 years ago (edited 2 years ago)

      yes, I felt the squeeze too, in trying to fit in a crisp talk. I was thinking that it will be better at around 25-30 mins. Will be editing this proposal with more content.

  • Anuj Gupta (@anujgupta82) 2 years ago

    I completely agree with Ananth, It should be a 40 min talk.
    @Ramanan : see if you can give away some notebooks to play with. Audience can then relate your talk to code

  • Ramanan Balakrishnan (@ramananbalakrishnan) Proposer 2 years ago

    After trying a mock rehearsal - looks like the full talk format is the way to go. The talk outline and slides have been updated with the new content.

Login with Twitter or Google to leave a comment