AI to generate novel images.

This post is regarding a recent kaggle competition that was quite different from other competitons, in that it was aimed at pushing the frontier of creative machine learning by developing techniques for generating novel images rather than more widely used tasks like regression, time-series prediction, or classification; the objective of this competition was to generating novel images of dogs using class of algorithms are called Generative adversarial networks (GANs)s, a class of machine learning systems that can generate data(images) that has never existed before– subset of generative models.
GANs can be considered as artists that create novel stuff from trite stuff by iteratively learning to do well on the task.

We were supposed to build models to create images of dogs. Only this time… there’s no ground truth data to predict. Here, we were instructed to submit the images and be scored based on how well those images are classified as fake dogs from the real ones.

Generative methods (in particular, GANs) are currently used in various places on Kaggle for data augmentation. Their potential is vast; they can learn to mimic any distribution of data across any domain: photographs, drawings, music, and prose. If successful, it will push the frontiers of generative image creation, but also enable to create more experiments across a variety of domains in the future.

In this competition I used Conditional GANs(cGANs) and worked majorly with keras to achieve promising results, although not the best. I achieved 238 place out of 927 teams! This was enthralling for me as this was my first ever kaggle competition in a field like GANs that is yet nuanced.

I started off with baseline work from other peers on kaggle platform, hence the reason why it is highly recommended to learn and enhance one’s skills of data science; thereafter I tried reading multiple reasearch papers on the subject of GAN’s, grasping the mechanics of the framework, and thereafter trying out multiple approaches & tweaking things here & there until I started getting good results.

Please check here for code & more information.