DARPA wants to give AI common sense using child psychology

Common sense is “perhaps the most significant barrier” between the focus of AI applications today, and the human-like systems we dream of.
2 November 2018 | 13 Shares

Dr Brian Pierce of DARPA. Source: Getty Images

As advanced as AI is today, the Pentagon’s Defense Advanced Research Projects Agency (AKA DARPA), thinks it lacks a crucial, and very human element: common sense.

Fresh off the back of announcing a US$2 billion, multi-year commitment to ‘third-wave’ AI tools, the government research body is launching itself into the development of reasoning and contextual awareness in the technology, which, it argues, has so far been lacking.

“Machine common sense has long been a critical—but missing—component of AI,” reads DARPA’s Machine Common Sense (MCS) proposal. “Recent advances in machine learning have resulted in exciting new capabilities, but machine reasoning remains narrow and highly specialized.”

While AI experts can train or program systems to respond to every situation, common sense— or the ability to understand and judge things that are common to all, without debate—has been elusive.

That includes, for example, an intuitive understanding of physics; knowledge of common facts an average adult possesses; and a basic understanding of human motives and behaviours. It helps us answer questions like “can an elephant fit through a doorway?”, and understand the sentiment “I saw the Grand Canyon flying to New York”.

While most humans share an understanding of this kind of knowledge, machines, being well, machines, lack it. It’s an abstract concept to articulate and program into machines, and according to DARPA, it’s “perhaps the most significant barrier” between the narrow focus of AI applications at present, and the human-like systems we dream of in the future.

With admittedly ominous ambitions to develop the technology for both commercial and defense purposes, DARPA’s plan for MCS is to accelerate the construction of intelligent systems capable of understanding their world, behaving reasonably in unforeseen situations, communicating ‘naturally’ with people, and learning from new experiences.

According to the research organization, there are four standout use cases for these advancements:

Sensemaking – For AI systems that need to analyze and interpret sensor or data input, a machine common sense service could help it understand and distinguish real-world scenes and situations.

Reasonableness –  A common sense service could give machines the ability to monitor and check the reasonableness of its AI system’s actions and decisions, particularly in ‘novel’ situations.

Collaboration – With human communication and understanding assuming a background of common sense, effective and advanced human-machine collaboration will rely on developments in this arena.

Transferable learning – Adapting to new situations could hinge on machines reusing common sense knowledge. This ability could provide a foundation to learning new domains without the need for additional specialized training or programming.

DARPA will attack research and development here by taking a two-pronged approach. One strategy will see teams developing systems for teaching machines through experience, mimicking the way babies grow to understand the world. An understanding of basic physics and spatial reasoning would be fostered through simulations of the physical world.

A second team, meanwhile, will analyze AI tools’ abilities within three cognitive milestones in the developmental psychology of children from birth to 18 months old, including prediction and expectation, experiential learning, and problem-solving.

“Today, machines lack contextual reasoning capabilities, and their training must cover every eventuality, which is not only costly but ultimately impossible,” said DARPA Director Steven Walker. “We want to explore how machines can acquire human-like communication and reasoning capabilities, with the ability to recognize new situations and environments and adapt to them.”

DARPA is looking to begin work in these areas by June 2019, with research estimated to take four years.