The researchers at IISc, Bangalore have submitted a paper for the Machine Learning Conference, NeurIPS 2020, that talks about how they taught the Generative Adversarial Networks (GANs AI) what not to learn and ignore. So how did they do it? How can this AI be a gamechanger for fighting against deepfakes?
Crux of the Matter
“The Art Of Knowing Is Knowing What To Ignore”
These words of the popular Persian poet, Rumi, inspired the IISc researchers to teach Generative Adversarial Networks (GANs AI) what to ignore and not to learn. i.e understand both examples and counterexamples.
What Are GANs?
They are machine learning frameworks that work on the zero-sum game theory i.e a mathematical representation of a situation wherein each participant’s gain or loss is exactly balanced by that of the other participants. If the total gains of the participants are added up and the total losses are subtracted, they will sum to zero.
What Consists Of A GAN?
A GAN consist of a generator, that generates candidates and a discriminator, that evaluates them. It is based on unsupervised learning i.e the network discovers and learns the patterns in input data, to generate new outputs on their own.
How Does The Rumi Framework Work?
The discriminator gives one of the following inputs:
1) Positives: samples from the target distribution. The generator has to learning this sample.
2) Negatives: samples that must be avoided.
3) Fakes: samples drawn from the generator.
Battle of frameworks begins as LS-GAN and AC-GAN are compared with RUMI-GAN framework via MNIST frameworks. Rumi-LSGAN generates sharper images consistently from the positive class, unlike the other two.
Where Are They Used?
- In completing the missing or damaged parts of a photographs and create photos of imaginary fashion models, saving the actual studio setup cost.
- Simulating and predicting scientific processes. Eg: Gravitational lensing for Dark matter research done in 2019.
- Video Game Modding, alteration to an existing video game, done by fans.
Classic Example of GAN
In a GAN, every time the input text is changed in the training set, a different image is generated with a single initial image taken as an input sample.
Understanding RUMI GAN can give us more insights in Deepfakes, which are manipulated videos and fabricated images that can be used to generate fake news by influencing everything from elections to stock markets.
Moreover, Deepfakes are often associated with GANs as even they study pictures and videos of a target person from multiple angles, and mimic his/her behavior and speech patterns.
What About Human Biases?
These networks and models aren’t representative of the whole population, when trained on unbalanced data with insufficient racial diversity, So even when RUMI framework can overcome the biases of the dataset, the biases of a data scientist still need to be tackled.
- The Music Genome Project is an effort to “capture the essence of music at the most fundamental level” using over 450 attributes to describe songs and a complex mathematical algorithm to organize them. The Music Genome Project was first conceived by Will Glaser and Tim Westergren in late 1999.
- The English interpretations of Rumi’s poetry by Coleman Barks have sold more than half a million copies worldwide, and Rumi is one of the most widely read poets in the United States. Recordings of Rumi poems have made it to the USA’s Billboard’s Top 20 list.
- Rumi and his mausoleum were depicted on the back-side of the 5000 Turkish lira banknotes of 1981–1994.
- Analytics India Magazine – IISc Researchers Teach AI To Be Ignorant
- Machine Learning Mastery – A Gentle Introduction to Generative Adversarial Networks (GANs)
- CNBC – What ‘deepfakes’ are and how they may be dangerous
- Noise Lab UCSD – “Deep Fakes” using Generative Adversarial Networks (GAN)
- ARXIV – Generative Adversarial Text to Image Synthesis