A new program created by scientists at Google DeepMind uses deep learning and an external memory to navigate the London Underground system. Although the task itself can be accomplished by a number of much simpler smartphone apps, the method used is considered to mark a breakthrough that will lead to new capabilities.
In the deep learning approach, an AI agent learns to accomplish tasks on its own, instead of being pre-programmed for each specific task by a human. Professor Geoff Hinton, now employed at Google, is considered to be the father of the deep learning method. He commented that the new program opens the door for a range of more complex applications for deep learning AI than ever thought possible.
“Until very recently, it was far from obvious how deep learning could be used to allow a system to acquire the algorithms needed for conscious deliberate reasoning,” according to Hinton.
Deep learning has enabled a number of breakthroughs already in AI language translation, image and speech recognition, and other tasks. A computer using deep learning even beat a top-ranked human player at the ancient strategy game Go. However, deep learning AI have continued to struggle at creating underlying strategies, for tasks such as navigation and interpreting the meaning of text.
By adding the use of external memory, the new program pushes the method forward on the front. External memory allows temporary storage of important pieces of information, to be retrieved when necessary. It reflects the human ability to plan a multi-step task such as following a recipe, or assembling a bookcase.
Development of Google’s new program was led by research scientist Alex Graves, who emphasized the incremental nature of the progress made by the program:
“I’m wary of saying now we have a machine that can reason,” Graves said. “We have something that has an improved memory – a different kind of memory that we believe is a necessary component of reasoning. It’s hard to draw a line in the sand.”
The work was published in the journal Nature.
In the study, the program was also asked questions relating to basic comprehension of text. It was given bits of a story, such as “John is in the playground. John picked up the football,” and then asked where the football was. In answering correctly 96 percent of the time, the program broke ahead on a task that most current AI programs are not able to perform successfully.
This may indicate that personal assistant systems like Apple’s Siri may soon be replaced by more complex programs with the ability to, for example, answer complex questions by searching the internet and carry out instructions on their own.