AI that can make anyone dance
This is just the beginning of what’s possible with Generative Adversarial Networks (GANS).
In the midst of this robotic revolution, now and then some real game changing technologies emerge that accelerate the pace of innovation. Introduction of Generative Adversarial Networks (GANs) into the wild world of AI is an example of a game changer. Ever since, several flavors of GANs have been introduced by various AI researchers to solve for many applications with great results.
For instance, researchers from UC Berkeley published an amazing work on motion re-targeting that can make fake videos showing amateurs dance like a professional dancer. In this work, various custom GANs were used along with other video generation and smoothening techniques. Check it out!
At a very high level, GANs involve two algorithms a “generator” and a “discriminator” pitting against each other in order to improve the abilities of the overall system. Keep reading to learn more about the introduction of GANs technology, recent uses of this AI, insights into the future, and a brief overview of key industry terms.
What is the “Do as I do” AI Project?
Venturing into a new era of GANS, researchers constructed a system that uses deep learning to transpose dance moves to amateurs with no advanced skills from a source video.
How does this process work?
In the research paper, the authors state:
“With our framework, we create a variety of videos, enabling untrained amateurs to spin and twirl like ballerinas, perform martial arts kicks or dance as vibrantly as pop stars.”
Where Did Generative Adversarial Networks Originate?
What started as a night out with fellow students in 2014 ended with the development of a whole new avenue that would shape the future of artificial intelligence. While he celebrated the recent graduation of a fellow doctoral graduate, Ian Goodfellow was asked by his peer to help him with a project that would enable a computer to create photos independently.
At that time, technology existed that used neural networks, or algorithms that were modeled similarly to neurons in the brain, as “generative” machines able to create plausible data independently. However, the results of this process were very primitive and prone to errors such as the blurring of faces or the omission of ears. In order to find a solution for the project the other students were working on, massive amounts of number crunching would be required, which would not work.
Goodfellow returned home and pondered on this challenge which led to an idea. He wondered what would happen if you pitted two neural network systems against each other. In the late hours he gave it a shot and coded his software. It worked the very first time. Thus, the concept of Generative Adversarial Networks (GANs) was born.
Taking these methods a step further, the “do as I do” motion transfer system performs video generation while preserving intricate details, i.e. the subjects’ faces. Recent advancements in image generation and general image mapping frameworks have laid the groundwork and enabled the authors to learn a mapping from pose to target subject.
Generative adversarial -networks have emerged in recent years and function by approximating generative models. Many purposes exist for GANs, especially image generation because of their remarkable ability to replicate detailed, high-quality images.
Differentiating Between AI, Deep Learning & Machine Learning
Artificial Intelligence (AI)
Artificial intelligence is the general category which all of these terms belong to. In a nutshell, AI is a reference to any intelligence performed by a machine that results in an optimal or suboptimal resolution to a problem.
Deep Learning: The Human Connection
Inspired by the function and structure of the brain (specifically the networks of interconnected neurons) – deep learning features algorithms that mimic the biological structure of the brain.
Machine Learning: Cognition Signs
Machine learning is the use of algorithms to interpret data, draw conclusions from it, then go on to make predictions or determinations about something in the world.
Future Implications of AI, Machine Learning and Deep Learning
Room for Improvement
Upon observation, you will notice that the subject dancer isn’t perfect. You will catch some waviness and choppy motions. Nonetheless, “Everyone Can Dance” is a shining example of what modern AI is capable of. Some fine tuning and a little time behind this development will render methods indistinguishable from the real subjects, which leads to the next point.
Cause for Concern
Obvious scenarios render the concern of GANs and motion transfer technology being used to implicate people in false scenarios. Reddit recently cracked down on deepfake pornography and the political sphere is concerned about false information spreading via faked videos. Furthermore, down to a very fundamental level, endless scenarios exist as to how these methods can cause interpersonal conflicts.
Many applications of machine learning have been facilitated by the emergence of deep learning methods. Collectively, these improvements grow the AI field. The technological ability to break down tasks into ways that utilize various machine assists will facilitate future developments, including:
- Driverless cars
- Better preventative healthcare
- Robot driven surgical procedures
How can GANS and motion re-targeting be exploited to solve some of the healthcare problems? For example: Can this technology be transferred to perform motion re-targeting from an expert surgeon to a novice surgeon?
We believe the technologies needed to codify an expert surgeon’s proficiency into an AI system are here and we will see them in action soon. Such tech will empower Novice surgeons and provide care at the same level as a skilled expert.
Needless to say that AI spectrum is evolving at lightning speed and there’s no telling how far this is going to take us and how fast. Remember this is only the beginning. We will close this piece with the following quote from Andrew Ng, Founder of deeplearning.ai and Coursera Deep Learning Specialization, Course 5:
“Deep Learning is a superpower. With it you can make a computer see, synthesize novel art, translate languages, render a medical diagnosis, or build pieces of a car that can drive itself. If that isn’t a superpower, I don’t know what is.”