Google’s DeepMind has revealed a radical new research project designed to give AI’s an imagination.
The breakthrough means that systems will be able to think about their actions, and undertake ‘deliberate reasoning.’
The radical system uses an internal ‘imagination encoder’ that helps the AI decide what are and what aren’t useful predictions about its environment.
‘Imagining the consequences of your actions before you take them is a powerful tool of human cognition,’ DeepMind said.
‘When placing a glass on the edge of a table, for example, we will likely pause to consider how stable it is and whether it might fall.
‘On the basis of that imagined consequence we might readjust the glass to prevent it from falling and breaking.’
This form of deliberative reasoning is essentially ‘imagination’, and is a distinctly human ability and is a crucial tool in our everyday lives, according to DeepMind.
‘If our algorithms are to develop equally sophisticated behaviours, they too must have the capability to ‘imagine’ and reason about the future,’ it says.
DeepMind says the work builds on its AlphaGo project, and hailed its ‘tremendous results’
AlphaGo uses an ‘internal model’ to analyse how actions lead to future outcomes in order to to reason and plan.
‘These internal models work so well because environments like Go are ‘perfect’ – they have clearly defined rules which allow outcomes to be predicted very accurately in almost every circumstance,’ DeepMind says.
‘But the real world is complex, rules are not so clearly defined and unpredictable problems often arise.
‘Even for the most intelligent agents, imagining in these complex environments is a long and costly process.’
The team previously revealed its work teaching its AI to walk.
DeepMind researchers trained a number of simulated bodies, including a headless ‘walker,’ a four-legged ‘ant,’ and a 3D humanoid, to learn more complex behaviours as they carry out different locomotion tasks.
The results, while comical, show how these systems can learn to improve their own techniques as they interact with the different environments, eventually allowing them to run, jump, crouch and turn as needed.
In a new paper published to arXiv, researchers from Google’s DeepMind explain how simple reward functions can lead to ‘rich and robust behaviors,’ given a complex environment to learn in.
The researchers set their simulations up against several obstacles, from hurdles to wall slalom, to help the AI characters teach themselves the best ways to get from one point to another.
And, footage from the study offers a hilarious look into the trial-and-error process.
As the characters run around throughout each simulated environment, they almost seem intoxicated as they flail, fall, collide with walls, and appear to trip over their own feet.
But, over time, they become far more successful in their efforts to navigate the different types of terrain.
As the team explains in the paper, the environments presented to the simulated bodies are of varying levels of difficulty, providing a setting in which they must come up with locomotion skills of increasing complexity to overcome the challenges.
‘In this work, we explore how sophisticated behaviors can emerge from scratch from the body interacting with the environment using only simple high-level objectives, such as moving forward without falling,’ researchers explain in a new post on the DeepMind blog.
‘Specifically, we trained agents with a variety of simulated bodies to make progress across diverse terrains, which require jumping, turning, and crouching.
‘The results show our agents develop these complex skills without receiving specific instructions, an approach that can be applied to train our systems for multiple, distinct simulated bodies.’
The approach relies on a reinforcement learning algorithm, developed using components from several recent deep learning systems.
According to the researchers, this type of method could help AI systems to achieve flexible and natural behaviours which can grow as they’re exposed to different situations.
Sophisticated motor control is a ‘hallmark of physical intelligence,’ the researchers write in the blog, allowing for everything from a monkey’s controlled movements through the trees to the complex navigation of a football player on the field.
And, as artificial intelligence continues to improve, they explain, these capabilities could soon allow computers to take on more complicated tasks.