• John Ennis

Eye on AI - August 6th, 2021

Welcome to Aigora's "Eye on AI" series, where we round up exciting news at the intersection of consumer science and artificial intelligence!

This week, we’ll be looking in detail at XLand, the new 3D world developed by DeepMind, which acts as an endless virtual playground for AI in which systems learn new skills through endless experimentation, similar to how children learn through play.


Enjoy!


New Virtual Playground Allows AI Systems to Train Themselves



The virtual AI world XLand, as outlined in the recent MIT Tech Review article “An endlessly changing playground teaches AIs how to multitask,” looks just like a 3D puzzle video game. Arenas are filled with brightly-colored shapes. Objects can be moved and manipulated. There even appear to be players moving about, or shapes acting as players, endlessly pushing objects. It reminds me of those old computer simulations on arcade games, the ones that played on repeat until a human took the controls and pressed the ‘start’ button. Yet unlike those arcade games, the videos displayed on XLand are not simulations. They’re AI systems being trained through ‘play’ in an endless arena, sensing color and using available objects to solve constantly evolving puzzles.


“Instead of developing just the skills needed to solve a particular task, the AIs learn to experiment and explore, picking up skills they then use to succeed in tasks they’ve never seen before,” writes MIT Tech Review contributor Will Heaven. “It is a small step toward general intelligence.”

XLand raining evolves from simple single-player games, such as finding or moving a particular object, to more complex multiplayer games like capture the flag, where AIs work as a team. Unlike other AI training systems, XLand doesn’t train AI participants in one particular task. Instead, it trains them to develop their own skills through reinforcement learning to improve by trial and error, similar to how children learn through play.


Why Training AI Through Play Could Lead to General Intelligence



If the idea of aimless play sounds childish, that’s because it is. Children play all the time. So too do wise adults. It’s an essential part of human learning, and what Einstein called “the highest form of research.” Training AI systems through play teaches them essential skills they’ll need to perform future. Children learn collaboration, social skills, empathy, and imagination through play –– imagination, as we noted in a recent post, is a particularly difficult skill to teach AI. The more AI systems play, the more skills they learn. By training in XLand, AI systems are able to play virtually without end.


“Some of DeepMind’s XLand AIs played 700,000 different games in 4,000 different worlds, encountering 3.4 million unique tasks in total,” continues Heaven. “Instead of learning the best thing to do in each situation, which is what most existing reinforcement-learning AIs do, the players learned to experiment… AIs that learned to experiment had an advantage in most tasks, even ones that they had not seen before.”

How might this kind of ‘play’ training lead AI to general intelligence? As we discussed in a former Eye on AI, one of the biggest impediments to AI becoming truly intelligent is its inability to understand the concepts of ‘same’ and ‘different.’ Children, even animals, are able to understand these concepts intuitively. Not so much for AI systems. By training AI systems through constant play rather than at one specific task at a time, where trial and error data is collected and retained over time, it’s not too much of a stretch to see how AI might soon be able to grasp these concepts in the not-so-distant future. This is not a certainty, of course. But one can imagine.


Other News



That's it for now. If you'd like to receive email updates from Aigora, including weekly video recaps of our blog activity, click on the button below to join our email list. Thanks for stopping by!