DeepMind creates 'imaginative' AI that can create and plan

DeepMind creates 'imaginative' AI that can create and plan

BY JAMES WALKER JUL 27, 2017

Google-owned DeepMind has announced an AI agent that is capable of "imagining" things and planning how to complete future tasks. The development brings fully autonomous AI a step closer by addressing one of the major shortcomings of current systems.

Computers are excellent problem solvers that can perform calculations at rates far in excess of the human brain. However, humans retain the upper hand in creativity and imagination. We can reason with ourselves, develop plans and think of abstract concepts that can't be defined.

In a blog post this week, DeepMind said it has been able to develop an AI that can "imagine" and "reason about" the future. The company added it has seen "tremendous results" with the system by giving AI agents the ability to interpret their internal simulations.

Handing the agent introspection abilities gives it the ability of questioning its own actions, in the same way humans do. This leads directly to the ability to learn strategies and make plans. The result is improved adaptability to new scenarios which can't necessarily be solved using logic alone.

The AI's reward network is programmed to recognise efficiency as well as logical performance. Contemplating the outcome of completing a specific action could allow the AI to recognise inefficiencies in its plan that it would otherwise ignore. It generates an interpretation of the scenario that more closely mimics a human approach.

DeepMind illustrated the significance of the development using a simple example. If a human is tasked with moving a glass, they'll place it away from the edge of a table. This ensures it's stable and won't accidentally fall on the floor. An AI could interpret the task differently. It may leave the glass closer to the edge so it can quickly pick it up later, without thinking of the potential consequences.

DeepMind's breakthrough development allows the agent to consider the possibility that the glass may fall. The agent would interpret the scenario similarly to a human, placing the glass in the centre of the table. DeepMind's also seen bots succeed in using their imagination when playing several puzzle games.

Implementing imagination in artificial intelligence has previously proved to be a challenge. Modern AI still tackles problems using a predominantly logic-driven approach. This leads to exceptional problem-solving abilities but doesn't offer an accurate representation of human thought.

Last week, DeepMind founder Dennis Hassabis explained the need for AI agents to possess an imagination. He called for greater cooperation between AI and neuroscience researchers, noting that artificial intelligence should be modelled on how the human brain operates.

While we're still a long way from machines with human levels of intelligence, DeepMind's new development is nonetheless significant. However, the company warned that its tests were conducted in closed "perfect" environments and aren't representative of real world performance. It joined the calls for more research in the area, describing the need for AI imagination as a "must" for the future.

"If our algorithms are to develop equally sophisticated behaviours, they too must have the capability to ‘imagine’ and reason about the future. Beyond that they must be able to construct a plan using this knowledge," DeepMind said. "But the real world is complex, rules are not so clearly defined and unpredictable problems often arise. Being able to deal with imperfect models and learning to adapt a planning strategy to current state are important research questions."



Comments

Popular posts from this blog

High-speed Hyperloop track ready for first trial run in Nevada

How the Fed went from lender of last resort to destroyer of American wealth

Apple leaps into AI research with improved simulated + unsupervised learning