AI Monthly Digest #9 – the double-edged sword of modern technology

This edition is all about AI morality-related themes, with a slight tinge of Talking Heads and Modern Talking. Earlier this year, deepsense.ai highlighted AI morality and transparency as one of 2019’s dominant AI trends. May bore out our thesis, especially as it relates to potential misuse and malicious intent.  At the same time, though, AI provides unique chances to support entertainment and education, as well as deliver new business cases.

A bigger version of GPT-2 released to the public

Open-AI has recently shown the GPT-2 model has set a new gold standard for natural language processing. Following the acclaimed success of the model, OpenAI opted not to make it public due to the risk of malicious usage, particularly to produce spam and fake news at no cost. This sparked an uproar. The industry good practice is to release AI research work as open-source software, so other researchers can push the boundaries further without having to repeat all the work done earlier from scratch. In other words – OpenAI threw up a major hurdle to NLP-model development by keeping GPT-2 under wraps. To support the scientific side of the equation while reducing the malicious threat, OpenAI releases some smaller-scale models to the public. The model it recently released operates on 345M parameters, while the best original model consists of 1.5B parameters. Every parameter can be seen as a virtual neuron inside a neural network, so OpenAI is basically reducing the brain it designed. The original network was released to OpenAI partners currently working on malice-proofing the system. The first independent applications of the downscaled network are already available at talktotransformer.com and onionbot headline generator. Why does it matter? OpenAI is currently facing a difficult choice between supporting the global development of AI and the fear of losing control over dangerous technology. In a world facing a potential avalanche of fake news and social media being used to perpetuate propaganda, building a system that writes coherent and convincing texts is undoubtedly dangerous. This case allows one to see all the AI-related issues in a nutshell, including the technology’s amazing potential, the real threat of misuse or malicious intent. So the case may serve as a precedent for future cases.

Talking heads unleashed

A group of scientists working for Samsung’s AI Center in Moscow and Skolkovo Institute of Science and Technology designed a model that can produce a convincing video of a talking head from a single image, such as a passport photo or even a painting. The model renders with consistency both the background and the head’s behavior. Most impressively, the model builds a convincing video of a talking head from even a single image of the frame. The solution is searching for a similar face that was analyzed and extracts facial features including a nose, chin, mouth and eyes. The movement of those features is then applied on the image, as shown in the video. The results are undoubtedly impressive. Why does it matter? Yet another AI ethics-related issue, the talking-head technology poses the threat of deepfakes, images that show a person making statements that he or she would never make. This raises obvious questions about the malicious ways such technology could be used. On the other hand, when deepfakes are used for special effects in popular movies, no one seems to complain and critics even weigh in with their acclaim. Some of the better-known examples come from the Star Wars franchise, particularly Rogue One, which features Leia Organa wearing the face of a young Carrie Fisher. AI has also proved itself useful in promoting art. By leveraging this technology it is possible to deliver the talking head of Girl with a Pearl Earring or the Mona Lisa telling visitors from screens about a painting’s historical context – a great way to put more fun in art lessons for kids. Or just to have some fun seeing what a Stallone-faced Terminator would look like. Again, AI can be used for both good and evil ends. The ethics are up to the wielder of this double-edged sword.

Modern Talking – recreating the voice of Joe Rogan

Another example of deepfake-related technology is using AI to convincingly recreate Joe Rogan’s voice. The text-to-speech technology is not a new kid on the block, yet it is easy to spot due to the robotic and inhumanely calm style of speaking. Listening to automated text-to-speech was usually boring at best while delivering the unintentional comic effects of robotic speech, all in the absence of emotion or inflection. Dessa engineers have delivered a model that is not only transforming text to speech, but also recreating Joe Rogan’s style of speaking. Joe is a former MMA commentator who went on to become arguably the most popular podcaster in the world. Speaking with great emotion, heavily accenting and delivering power with every word, Rogan is hard to mistake. Or is he? The team released a quiz that challenges the listener to distinguish if a given sample comes from a real podcast or was AI-generated. The details can be found on Dessa’s blog. Why does it matter? Hearing a convincing imitation of a public personality’s voice is nearly as unsettling as watching a talking head talk. But the technology can be used for entertainment and educational purposes. For example, delivering a new Frank Sinatra single or presenting Winston Churchill’s comprehensive and detailed speech on reasons behind World War II. Again, the ethics are in the user’s hands, not in the tool. Despite that, and as we saw with OpenAI’s GPT-2 Natural Language Processing model, researchers have decided NOT to let the model go public.

Machine learning-powered translations increase trade by 10,9%

Researchers at Olin Business School at Washington University in St.Louis have found a direct connection between machine learning-powered translations and business efficiency. The study was conducted on e-Bay and shows that moderate improvement in the quality of language translation increased trade between countries on eBay by 10.9%. The study examined the trade between English speakers from the United States and their trade relations with countries speaking other languages in Europe, America and Asia. More on the research can be found on the Washington University of St.Louis website. Why does it matter? While there is no doubt that AI provides vital support for business, the evidence, while voluminous, remains largely anecdotal (sometimes called anec-data) with little quantitative research to back up the claim. Until the Olin study, which does provide hard and reliable data. Is justified true belief knowledge? That’s an entirely different question…

A practical approach to AI in Finland

AI Monthly Digest #5 presented a bit about a Finnish way of spreading the word about AI. Long story short: contrary to many approaches of building AI strategy in a top-down model, Finns have apparently decided to build AI-awareness as a grassroots movement. To support the strategy, the University of Helsinki has released a digital AI course on the foundations and basic principles of AI. It is available for free to everyone interested. Why does it matter? AI is gaining attention and the reactions are usually polarised – from fear of losing jobs and machine rebellion to arcadian visions of an automated future with no hunger or pain. The truth is no doubt far from either of those poles. Machine learning, deep learning and reinforcement learning are all built on certain technological foundations that are relatively easy to understand, including their strengths and limitations. The course provides good basic knowledge on these issues, which can do nothing but help our modern world.