Today, we’re excited to have Prof. Hao Li, CEO and Co-founder of Pinscreen, Assistant Professor of Computer Science at the University of Southern California and Director of the Vision and Graphics Lab at the USC Institute of Creative Technologies. He’ll be sharing his thoughts on resolving the Uncanny Valley Problem and the potential uses of human-like virtual avatars.
Wan Wei: Hello Hao Li, thank you for being with us here today. Can you tell us more about yourself and what you’re currently doing?
Hao Li: I’m a Professor at USC and the Director of the Vision & Graphics Lab of the USC Institute for Creative Technologies. I’m also the CEO of a start-up called Pinscreen.
The work that I do at the university is to research on all aspects of how to digitise humans using AI. At Pinscreen, we build products around consumer accessible 3D avatars to solve your everyday problems. For both, we’re advancing technologies that allow people to create their digital selves.
On Resolving the Uncanny Valley Problem
Wan Wei: What made you interested in this field of creating Human-like Artificial Intelligence or Human-like avatars?
Hao Li: There are a lot of things that are fascinating about it. Firstly, in terms of the technology itself, it’s a very, very challenging problem to solve because in some ways in our field, building faces that look realistic and overcoming the Uncanny Valley have always been the holy grail because we are very familiar with how we look like.
It’s also a very practical thing because,
Wan Wei: As you mentioned, overcoming the Uncanny Valley is the holy grail. How close do you think we are to resolving that effect?
Hao Li: I think we are on the right path of solving the uncanny valley problem. Traditionally, we always had difficulties overcoming the Uncanny Valley because of the way computer graphics works. We use different models that approximate reality, and when you put all these different models together, it becomes difficult to ensure that they can accurately reproduce something as sensitive as the human face. You would still be able to tell that synthetic faces are fake, which can be very off-putting.
Nowadays, using all these data-driven approaches, many of them using deep learning, that are trained by real photographs and real videos, you can generate something that looks completely authentic.
One thing that is funny is that you sometimes get a different kind of uncanny valley. If you look at some of the videos and images that we generate using deep generative models, teeth often look wrong. Yet, I think these are problems that we can fix and, in many ways, the classic Uncanny Valley problem will soon not exist anymore.
On the Potential Uses of Human-like Virtual Avatars
Wan Wei: As a general audience, human avatars are usually just characters in a video game. In your work on digitising humans and at Pinscreen, how difficult is it to create realistic virtual humans?
Hao Li: Not everyone knows how difficult it is to create digital humans. To many, the first thing they think of is something that looks very game-like. I think it depends a little bit on the audience that you ask.
The reality is, it’s still very difficult to recreate a virtual human that looks like real ones. The technology for doing that automatically isn’t entirely there yet. While we made a lot of things possible recently, it’s still quite difficult to create an avatar automatically that is indistinguishable from reality.
So that’s the first problem that we must solve. We need to be able to convert research prototypes into something that actually works in the real world, and in this case, it means that we need to take our digitisation algorithms outside of research
Wan Wei: So, what are some of the practical use cases you see for digital humans?
Hao Li: Some are very obvious. If you can create a digital version of you, you can play yourself inside a game. So just imagine if you play a [game] like GTA or Fortnite, then you can play with together with your friends or even celebrities in the game itself.
Another application that is very important is the field of communication. So, right now, we are still using video chats, and it’s not ideal. It doesn’t replace actual in-person meetings.
We can think of using our avatars to have an immersive version of yourself, to be able to replace a physical face to face meeting. You could wear a VR headset and suddenly, it’s as if you are talking in the same location even though you’re completely far away from each other.
Wan Wei: Your speciality is in creating virtual human avatars using AI. In the AI field in general, if you could choose only one area in AI to watch this year, what will you choose?
Hao Li: Personally, I would be really interested in looking into true AI. This should not be confused with the type of applied AI we are using in our field, which is a very powerful tool in a sub-area of machine learning called deep learning that people are applying widely. Deep learning can solve a lot of problems that were traditionally very difficult to solve such as pattern recognition, image synthesis, and image segmentation. However, it’s not actually true artificial intelligence as many of us may imagine. It doesn’t have any consciousness. It’s purely driven by data, and it’s doing an advanced form of pattern recognition.
True AI is often referred to as artificial general intelligence. Even though the first stage is about fooling a real human being and being able to conduct real conversations through a 3D avatar, there is a lot of work to be done in creating faithful behaviours. This would involve natural language processing, motion synthesis, and how to manage some level of memory and emotions.
Wan Wei: Thank you for your insights. Before we end, do you have any final comments?
Hao Li: I would like to recommend you try our app on Pinscreen. We have the best technology that creates 3D avatars from a picture, and in a few years, we will be able to create the perfect clone of yourself that can even behave and think like you.
Wan Wei: Thank you for your time and your insights Hao Li.
Hao Li: Thank You.