Replay at https://www.youtube.com/YouthAILab
Brief Summary:
Dr. Joelle Pineau is an associate professor and William Dawson Scholar at the School of Computer Science at McGill University in Montreal, Canada. She also leads Meta (Facebook)’s Fundamental AI Research lab. In the talk, Dr. Pineau discussed the main areas of AI research that will influence the development of the metaverse, and what this opens in terms of applications. In the talk, Dr. Pineau first talked about the early days of computer intelligence technologies, from programmatic AI to machine learning. She talked about how programmatic AI was replaced by machine learning as it was particularly effective for problems where it was difficult to write the rules of the game, such as brain scans. Dr. Pineau also explained the definition of “learning” in terms of AI technology. Prof. Pineau then presented neural networks, a modern technique to build prediction machines. She compared how an artificial neural network is formed using a similar structure to the human neuron, with mathematical equations that represent the process of the human brain. Ultimately, she introduced the idea of the metaverse, a technology that she thinks is the future form of the internet. She concluded her presentation with the conclusion that because there are such a wide set of choices that will be made using AI technology in the future, it is important to observe the direction of future technologies and utilize them for your own good.
Full Summary:
Dr. Joelle Pineau is an associate professor and William Dawson Scholar at the School of Computer Science at McGill University in Montreal, Canada. She also leads Meta (Facebook)’s Fundamental AI Research lab. She received an MSc and P.h.D. in Robotics from Carnegie Mellon University. Her research focuses on complex problems in robotics, health care, games, and conversational agents. Along with generating new medical treatment strategies, her team has used Deep learning to detect seizures and AI technology to analyze medical scans. As a credited AI expert, Dr. Pineau discussed the main areas of AI research that will influence the development of the metaverse, and what this opens in terms of applications.
In the talk, Dr. Pineau first talked about the early days of computer intelligence technologies, from programmatic AI to machine learning. She said that up to the 1990s a lot of people thought the best way to crack AI, meaning to build intelligent systems such as those to play chess, was really based on instructing the machine how to solve problems. Scientists would give very precise instructions in terms of the rules of the game and let the system search through a very wide space of solutions. While this technique was widely used during the period, it did not produce good results and the rise of machine learning replaced it as large enough datasets became available. Early usage of machine learning was particularly effective for problems where it was difficult to write the rules of the game, such as brain scans.
Dr. Pineau went further to explain the definition of “learning” in terms of AI technology. She introduced a class of methods called supervised learning. In a supervised learning system, a person would show the machines several examples of a particular topic and pair those examples with a concept. By letting the machines see lots of examples, they can learn some general properties of the subject it is being taught; in particular, learning to generalize to new pictures of the same type of object. This has been the predominant way we train machines from the 1990s to modern days.
Prof. Pineau then presented neural networks, a modern technique to build prediction machines. Neural networks that we build in our computers are actually inspired by the biological neural networks, she said. She compared the structure of biological neurons and artificial neurons; biological neurons receive information from other neurons to compute some information and decide whether to send off an impulse, and artificial neurons would perform similar functions, through mathematical equations that model the human neuron. Through layers of thousands of these artificial neurons, we can build artificial neural networks. These networks are trained by seeing many input-output pairs, as well as the implementation of “weights” (represented by theta) that are set to minimize the number of errors and can be learned by gradient descent algorithm.
Dr. Pineau also discussed the future of neural networks. She illustrated a timeline of how neural networks evolved throughout the years, pointing out the fact that the pace at which AI technology advanced actually accelerated, reaching the next level within just a few years. She then gave real-life examples of neural networks such as the smart wheeler system that was built a few years ago and devices that can be installed to assist disabled people. Prof. Pineau drew a rough estimate of the time it will take humanity to reach a higher level of AI technology through the usage of graphs. From AlexNet, the first artificial neural network, to AlphaGo Zero, one of the machine learning programs by Google Deepmind, Dr. Pineau shows the exponential curve of AI technology development to highlight the widening application of neural networks.
She then brought in self-supervised learning, the key to building a unified model which then, with a little bit of adaptation, can be used with many different tasks. In self-supervised learning, you get the image but no one tells you what the concept is. Through this training system, AI does not learn a specific concept designated to it; rather, it learns the representation itself, and the nature of the data. Because the machine is trained with millions of other images, it gets the sense of what would be a plausible way to fill in the image given all the other images. She said that it is surprising how well it reconstructs even if the majority of the image is blocked. It is the best hope we have right now of being able to build a unified model.
Dr. Pineau then raised the idea of the metaverse, which, in her words, is the next generation of the internet. The idea of the metaverse is to have a massively scaled and interoperable network of real-time rendered 3D virtual worlds that can be experienced synchronously and persistently by an effectively unlimited number of users with an individual sense of presence and with continuity of data, such as identity, history, entitlements, objects, communications, and payments. She said that she is excited with the technology because she believes it is going to be built by the upcoming generation, the audience, much more than by the previous generation. She drew attention to the life incorporated with metaverses, such as virtual workplaces and virtual entertainment spaces where people can be more creative as their imagination can be created in the virtual world visually. Dr. Pineau noted that it is important to decide the kinds of experiences that we want to have in this new space similar to the real world. She wrapped up her presentation by saying that because there are such a wide set of choices that will be made using AI technology in the future, it is important to observe the direction of future technologies and utilize them for your own good.
For previous Youth AI talks, visit https://www.youtube.com/YouthAILab
0 Comments