On Machine Intelligence and Knowledge

Source: Deep Learning on Medium

Once upon a time, humankind invented machinery for mathematics. Over time, we discovered that these machines could do more than count money or solve the physics of aiming artillery. We dreamed that our mental abilities were purely mathematical. Like gods, we might be able to create ourselves.

Today, we have invented software to run these machines that can recognize a Chihuahua and drive a car. How do these compare with our own intelligence? What do these devices really know? Most importantly, where do we go from here? Should we be afraid?

Some of the first achievements of AI were in a toy world made of abstract blocks symbolized as a set of numbers. Programmed rules could help decide if a block was supported by another block — these relations were returned as more symbols. Blocks World was very primitive, but it illustrated what this kind of machine was capable of. In hindsight, not much of any use at all.

Artificial intelligence today still operates entirely within toy worlds. Natural intelligence such as that inside a rat’s head is capable of being dropped off in a jungle and learning to survive. This end-to-end observation, cogitation, and reaction is the goal of Artificial General Intelligence — to achieve parity with human intelligence. Our software today, as thought-provoking as it is, does not come close.

Intelligence, whether embodied in a brain or machine, manifests as this end-to-end stimulus-response feedback loop with the real word. To differing degrees, a frog catching a fly with its tongue and the engineer catching a plane to get to a job interview both exhibit the intelligence they need to survive in the real world. (Yes, the engineer’s world is artificial, but no less real.)

The ability to react in these ways, in whatever manifestation, is what we call knowledge. The frog knows how to use the body it was born with to catch a fly with its tongue. I have neither the knowledge nor the body for that, although I can catch a plane. The brain, in its neural connections, and the deep learning AI, in its weight matrices, embody the know-how of how to react in this end-to-end manner.

Applied knowledge is what we call intelligence.

Today’s best and brightest AIs only capture a bit of that stimulus-response pathway of synaptic impulses between the real world and our actions to survive and prosper within it. The best image classifiers, for example, can perform better than humans, but only in an abstracted and simplified toy world where the visual appearance of some thing is captured in a frozen moment in time, and returned as a number signifying whether it is a cat, or a tumor, or a bicycle. Our self-driving cars, while so very useful, exist in a simplified environment of roads and obstacles both moving and stationary. Their action, their world of possibility, is limited to two numbers — acceleration and steering angle. Cars, or robots of any kind, cannot hunt, shop for groceries, or catch a train.

Conversational AIs come the closest to this end-to-end ability within a thin, simplified, yet unstructured and realistic (indeed, real) world, composed of the time sequence of sound waves within a room. Set loose on a low bandwidth world such as a telephone line, it’s able to survive and prosper in our world of making hair appointments or ordering Chinese food from a restaurant.

There is much opportunity for improvement. Our performant AIs can be linked together in an end-to-end fashion, and areas which need improvement can be identified. New research on how animals think can shape our directions and implementations. I’m particularly fascinated with models of intelligence that shake our ideas of how brains think, such as David Cox’s work on predictive modeling in the brain.

I’m truly excited and amazed by recent achievements of artificial intelligence and their brilliant creators. I do believe that someday we can build intelligence that has great advantages over our own, such as being wired into a network of servers, and operating (thinking) faster than chemistry allows. I’m convinced this is inevitable. Am I afraid of what can happen? A little bit.