Fake it until you make it ..Generative Adversarial Neural Nets to create and resolve deep fakes
What is the role of AI in the content creation arena? What are the societal and cultural aspects of big challenges like fake news and how AI can fix them?
Delip Rao, the vice president of research at the AI Foundation and creator of the Fake News Challenge shares his main views this super intersting podcast.
Here's 5 things to remember:
There's no silver bullet for solving the fake news problem.
Fake news is not a single monolithic problem, it is a combination of very complex multifaceted issues. One way to unbundle this hairy challenge is to combine Natural Language Processing modeling with a human-in-the-loop approach (e.g wisdom of the crowd, manual annotation, and human supervision).
Volume is among the most challenging issues in detecting non-original content and fake news.
A lot of news is simply a replica that is fueling the echo-chamber phenomenon. While Machine Learning can detect plain replicas, the same techniques have hard time to understand if a news was forged or it contains factual, truthful information. At the moment manual curation is the only trusted solution
Build a dataset → create a competition→ build a community → solve a big problem.
This is an increasingly common, and surprisingly effective strategy for attacking research challenges. You create a novel dataset that hasn't seen before, you raise interest around it through challenges in platforms such as Kaggle, DrivenData, Innocentive, Crowdai, Numerai or Tianchi. Smart people will respond with solutions!
Beyond fake news: Deep fakes are now polluting all three of the dominant modalities text, audio, and imagery.
With GANs (Generative Adversarial Neural Networks), an AI technique recently invented in Stanford, you can achieve surprisingly good results when modifying faces, swapping facial attributes, or even creating a new face from scratch. The results are creepy and very realistic!
5. Fake news Detection vs. Generation algorithms, the battle is intensifying.
Computers are now able to fake not only articles, but also videos and audio with your voice. When it comes to audio, computers can artificially generate voice, but they are also very effective at detecting it. For instance, the most common attack (e.g. replay attack) can now be detected with extremely high accuracy. So the computers create the problem of fake content but they can also solve it!
..buy Natural Language Processing with PyTorch which seems like an interesting starting point if you are developing Deep Learning NLP models
..listen and compare the quality of voices generated with traditional approaches and state of the art approaches created with Deep Learning
..read about the Fake News Challenge, where 100 teams and 1000 competitors participated in foundation experiments. -AI.foundation*