What do the robots know and when did they know it? But more importantly, how did they learn it? Technology is improving and advancing at a blistering pace and the implications for AI, advances in robotics, and more, depend heavily upon learning. Michael Littman, Professor of Computer Science at Brown University will delve into a broad discussion of the multiple types of learning as they relate to multi-layered neural networks.
Professor Littman provides examples of the various kinds of learning that are available and how they are suited to various tasks. The Brown University professor discusses machine learning, which applies to the creation of systems that use data and AI to improve targeted areas of functioning. We'll also gain insight into supervised learning, which is learning based on feedback, essentially correcting a response via given feedback. This process can provide advanced teaching in an AI environment via various inputs such as layers, such that the learning system can match input to the desired output. For example, as we consider a particular image, individual layers will match data until eventually, in aggregate, that image can be classified and thus the input leads to a conclusion of what that image actually is.
Further, Mr. Littman discusses another specialized type of learning that can be applied known as reinforcement learning, which would simply allow an AI network to make selections on its own, then at the end, it would be given reinforcement. Reinforcement would inform the network as to whether it has been successful or whether it has failed, which would allow the network to learn from either and advance.
Additionally, Mr. Littman will explain how algorithms can map out narrow pieces of multi-dimensional space within an entire network to gain clues about what could be improved. With technological advances that allow for more data to be gathered, along with accelerated computer processing speeds and better algorithms, networks can be set up and configured to produce improved results for training data.
Linking the future of learning to our cultural past, Littman provides an interesting overview of how various types of network learning were applied to early Atari video games. The learning allowed for a network to equal or surpass human scores. But this research certainly goes well beyond teaching a network to win an old school video game; the applications for this learning are directly applicable to advanced robotics. With applied training and learning, extremely advanced AI is more than just a researcher's dream in some tech laboratory—it's coming.