Nicu Sebe, Head of Artificial Intelligence at Humans.ai, chief academic figure and resident rocker, offered from his expertise as an AI and ML professor some thought-provoking predictions concerning what will happen with AI in 2023, highlighting what the future holds in store for humanity’s most powerful technology.
With a wealth of academic knowledge under his belt, Nicu Sebe offered a peek into what is happening in the academic environment concerning artificial intelligence, stating that the work and research performed in labs will significantly impact how technology will evolve in 2023. So, without further ado, these are our Head of AI’s tech predictions for 2023.
Before focusing on what 2023 may have in store for the AI tech scene, a brief introduction is in order. Nicu Sebe is a professor of Computer Science at the University of Trento, Italy, where he leads the research in the areas of multimedia information retrieval and human-computer interaction in computer vision applications. For a year, Nicu Sebe has been part of the Humans.ai deep tech company, where he is working on generative computer vision and multimedia, trying to develop more efficient ways to generate content and animate faces and voices.
Humans’ Head of AI shared that things have developed at a rapid pace in the last couple of years, especially in the field of AI-generated media.
“In the beginning, we were taking 2D forms and modeling them. For example, we were unable to animate very well 3D motion. The next step was to have some kind of 3D models and animate them. What started about 2 years ago and is now very popular is to have these neural diffusion fields. It’s something that rests at the border between computer vision and computer graphics in which you generate a few camera positions, and then you do some sort of seamless interpolation between different views to generate a 3D View of the scene”.
As Humans.ai’s expert underlined, things are changing at a rapid pace in the world of AI, each year bringing new concepts and methods of doing things.
Our expert shares that what Humans.ai is looking towards, from a technological standpoint, is to adopt an object-agnostic approach. Simply put, this means that the company is exploring how it can expand its ability to animate faces and the upper torso of a person to also include full-body animation, which is far more complex.
The professor shares that nowadays, many companies, including Humans.ai, have taken the technology and concepts developed by AI researchers and implemented them in multiple use cases. He predicts that in the future, companies will focus on more complex facial animations like micro-expressions and full-body animation. In his opinion, this can be achieved by focusing on a multimodal approach that would enhance the experience and the end result by including different modalities that will enable people to use text to speak in multiple languages or generate animations.
“I think, in many cases, you would want to have an alter ego. So, you could use a digital copy of yourself to do things for yourself. Of course, this brings up some clear problems because, of course, you need to make sure that it is yourself or if it is the owner of the content that you are animating.”
Underlined Nicu, while drawing attention to some of the things that companies and people who use this type of technology need to be cautious of.
Content generation AI use cases peaked in popularity in 2022 with applications like DALL-E and ChatGPT. Even so, these types of applications are still in their very early stages, requiring more training and data to help them generate more accurate images, in the case of DALL-E, and formulate more relevant responses and conversations, for ChatGPT.
A popular example that demonstrates this is DALL-E’s majestic rendition of salmon. Many people who played around with the software and tried to generate something like “salmon in the river” or “salmon swimming upstream” did receive a picture of a salmon swimming in the river, the problem is that the salmon in question were prime cuts of salmon fillet that were majestically swimming in the river. Even so, one cannot deny the good laughs generated by this imagery. Although, it goes to show that we aren’t quite near perfect implementation of generative AI.
As Humans’ Head of AI explained:
“This is a clear example of which AI is kind of blind because it just learns from data. It’s seen a lot of salmon fillet, and it believes that is what a salmon looks like. So, one of the bottlenecks is data. What these guys are doing is based on a huge amount of data”.
Nicu Sebe points out that in 2023 after additional data is fed to these types of systems and more training is performed, we will witness an increase in the quality of generative AI. But this technology can be used in more complex ways. For example, you can have a seamless interpolation between different paintings, so you have something like the Girl with a Pearl Earring (an oil painting by Dutch Golden Age painter Johannes Vermeer) and extend the room she is in so you can go outside to expand the universe she exists in and create a more immersive experience for the beholder.
“All this AI technology can be used to do the interpolation. It’s working really well. Of course, one of the main things is that there is no ground truth for this. You can’t tell that one representation is better than another one. It’s just in the eye of the beholder. You can like it, or you don’t. I think that’s why, to me, the most important aspect is that the human is the one who is consuming and the one who decides what he likes. I think this is really the next frontier of generative AI in general”, concludes the professor while offering his view on generative AI.
An unexpected side effect of new opportunities for creativity is that it reshapes how people view and use technology. Usually, when new tools are invented, people get creative and invent new ways to use them, which can take creativity in totally unexplored areas.
One of the main focuses in the field of artificial intelligence is called trustworthy AI. An interesting phenomenon that goes often unnoticed in the world of deep fakes is the parallel movement which works to detect and decodify AI-generated deep fakes. AI generative technology has evolved to a point in which the images they produce cannot be detected by the naked human eye, which by itself raises serious concerns.
Commenting on this issue, Nicu Sebe unveiled for the first time that he is part of a European initiative that aims to help combat the growing deep fake threat.
“I am now part of a European project called AI for Trust, which focuses on detecting deep fakes. Our intuition was that since we know how to make deep fakes, we should also know how to detect them.”
The AI expert also highlights the fact that a concept that has recently emerged in the field of AI can help boost the levels of trust. This concept is called unlearning, and it works by making AI forget certain unwanted patterns. For example, you train an AI to learn something with data from Romania, but it manages to pick up some cultural biases as a side effect. To circumvent any future problems that may appear if you use the same AI in another country, like Nederlands, you may want to have that particular AI unlearn some patterns. In a sense, to eliminate the Romanianess from it and make it learn something on top of that information.
Another buzzword besides trustworthy AI is explainable AI. Most deep learning models are like black boxes. So, what scientists and researchers try to dissect these modelsto understand parts of it, which layers are doing what and to try to control them, to inject information into particular layers to control what’s going on in the output. This is another trend that people are working on nowadays.
Artificial intelligence models require vast quantities of data necessary for training and fine-tuning. This high data demand is met thanks to the millions of users from all over the globe who generate a constant stream of new information each day. The catch is that not all data is useful, secure or ethically harvested and used.
Blockchain can address some of the major pain points that revolve around how data is handled and secured. Furthermore, due to its unique properties, this technology can help ensure AI’s fair and ethical use. Humans.ai is already doing groundbreaking work in this area by using blockchain to give people control over how artificial intelligence is used.
Companies are giving students access to their data and tech resources to help grow the researchers of tomorrow. Philips, Qualcomm, TomTom and Microsoft have shared labs with the university of Amsterdam. A joint lab is where companies put in the data and the use cases.
In this type of lab, there are not only researchers but also students who can get access to the data and play around with it. These types of collaboration are beneficial to all the parties involved. This way, academic researchers get access to real data and real problems. Instead of playing with toy applications, they get to solve real-world problems. Also, having access to the industry’s resources, like near-unlimited computing power, gives researchers a more in-depth perspective on different possibilities and ways of using data to understand and decode things.
A beneficial side effect of these collaborations between tech companies and academia is that it paves the way towards industrial PhDs where the goal isn’t necessarily to focus on only publishing research papers but also to develop real-life algorithms that can be applied directly in the industry.
Humans’ Head of AI shared that in November 2022, he submitted a proposal focused on this area alongside 30 European partners.
“Our goal was to look at sustainability from different levels. One area is global sustainability, which can cover multiple areas like how different materials are used and processed, how to use AI for efficiency in the energetical sector, how you can use AI to study climate changes, and so on. Then, you could go towards the society level and use AI to try to understand, for example, things like fluid democracy and different types of biases”, explains Nicu Sebe.
AI will maintain its ascending trend in 2023, marking a new degree of maturity for the technology as well as the emergence and implementation of new and exciting use cases that will have a more practical approach focused outside the boundaries of entertainment.
The new year will also mark the beginning of the expansion of Humans’ approach to AI, embracing a multimodal approach to the technology to supercharge user experience and immersion. In the words of our high-class expert:
“Visual information is important, and it is what makes us human, but we also have other senses. As a company, we need to focus on multiple modalities with techniques like text-to-voice, voice-to-text, text-to-image, etc. All these types of interplays are something that we will address.” Nicu Sebe