Data Science Conversations

This is Part Two of our conversation about Deep Fakes with two experts in their respective fields.

We talk to Dr Eileen Culloty of the Institute for Future Media and Journalism at Dublin City University and Dr Stephane Lathuiliere of Telecom Paris.

EPISODE SUMMARY:

• The implications of disinformation on the media
• The role of fact checking
• How to deal with Deep fakes
• What Adobe and Microsoft are doing about Deep Fakes
• The future role of GAN’s in detecting Deep Fakes

Resources:

Video of First Order Motion Model For Video Animation:

 https://www.youtube.com/watch?v=u-0cQ-grXBQ&ab_channel=AliaksandrSiarohin

 PROVENANCE program: https://fujomedia.eu/provenance/

This is Part one of our conversation about Deep Fakes with two experts in their respective fields.

We talk to Dr Eileen Culloty of the Institute for Future Media and Journalism at Dublin City University and Dr Stephane Lathuiliere of  Telecom Paris.

Stephane reveals what is possible and what is not possible technically with current Deep Fakes Technology.

Eileen helps us cut through the hype about Deep Fakes and tells us about their real world social and political impact.

EPISODE SUMMARY:

• Short History of Media Manipulation
• The breakthroughs in Deep Learning enabling current Deep Fake Technology.
• The role of increased data availability in generating Deep Fakes.
• Why Cheap fakes are still a bigger problem than Deep Fakes.
• How the First Order Motion Model has advanced the field of Image Animation.
• Positive use cases of Deep Fake Technology.
• The Future of Image Animation/Deep Fake Technology.
• Challenges for media and Journalism in the age of Deep Fake Technology.
• Societal impact of Disinformation and fake news content.
• Deep Fakes V Cheap Fakes during COVID Pandemic.

RESOURCES:

Video of First Order Motion Model For Video Animation:

 https://www.youtube.com/watch?v=u-0cQ-grXBQ&ab_channel=AliaksandrSiarohin

PROVENANCE program:  https://fujomedia.eu/provenance/

This is Part 2 of our conversation with Professor Philipp Koehn of Johns Hopkins University.  Professor Koehn is one of the world’s leading experts in the field of Machine Translation & NLP.  

In this episode we delve into commercial applications of machine translation, open source tools available and also take a look into what to expect in the field in the future.

Episode Summary:

• Typical datasets used for training models

• The role of infrastructure and technology in Machine Translation

• How the academic research in Machine Translation has manifested into industry applications

• Overview of what’s available in Open source tools for Machine Translation

• The Future of Machine Translation and can it pass a Turing test

Resources:

Philipp Koehn latest book - Neural Machine Translation - Amazon link: 

https://www.amazon.com/Neural-Machine-Translation-Philipp-Koehn/dp/1108497322

Omniscien Technologies - Leading Enterprise Provider of machine translation services:

https://omniscien.com/

Open Source tools:

- Fairseq https://fairseq.readthedocs.io/en/latest/

- Marian https://marian-nmt.github.io/

- OpenNMT https://opennmt.net/

- Sockeye https://awslabs.github.io/sockeye/

Translated texts (parallel data) for training:

- OPUS http://opus.nlpl.eu/

- Paracrawl https://paracrawl.eu/

Two papers mentioned about excessive use of computing power to train NLP models:

- GPT-3 https://arxiv.org/abs/2005.14165

- Roberta https://arxiv.org/abs/1907.11692

Professor Philipp Koehn of Johns Hopkins University discusses the evolution of machine translation and the fundamentals for using Neural Networks to deliver Machine translation.

Episode Summary:

• Philipp Koehn Bio
• What is Machine Translation?
• Adequacy & Fluency
• How to Quantify the performance of Machine Translation models
• The Transition from Statistical approaches to using Neural Networks for translation
• Validating Outputs of models
• What can go wrong with Machine Translation?

Resources:

 Philipp Koehn latest book - Neural Machine Translation - Amazon link: 

https://www.amazon.com/Neural-Machine-Translation-Philipp-Koehn/dp/1108497322

Your co-Hosts Damien Deighan and Philipp Diesinger discuss the Data Science conversations Podcast concept. You will find out what to expect from the show in the coming weeks.