Rebooting AI: Building Artificial Intelligence We Can Trust

learning but poor at compositionality and the construction of cognitive models; the other, classical AI, incorporates compositionality and the construction of cognitive models, but is mediocre at best at learning.

In classical AI, researchers would typically encode by hand the knowledge the AI would need to carry out a particular task, and then write computer programs that leveraged that knowledge, applying it to various cognitive challenges,

could discover a new use for a turkey baster. On a well-controlled factory floor, situational awareness can similarly be a relatively manageable problem, limited to questions like “Is there an obstacle here?” and “Is the conveyor belt running?” In the home, on the other hand, situations, and the risks, rewards, and opportunities that go with them,

Causality, broadly interpreted, includes any kind of knowledge about how the world changes over time.*1

simple object-recognition systems fall far short of understanding the relations between objects in a scene:

You can often find out whether something is inside your pocket by feeling the outside of the pocket. This is an example of how different types of knowledge may be combined. Here knowledge about how different objects (hands, pockets, wallets) interact with one another is combined with knowledge about how the senses work.

What we need is not just “deeper” learning, in the sense of having more layers in a neural network, but deeper understanding. We need systems that can truly reason about the complex interplay of entities that causally relate to one another in an ever-changing world.

guaranteed in “deeper” networks with more than two layers, it was possible to build a system that could produce results that were often good enough, opportunistically climbing up the mountain by taking small steps of the right sort, using a technique called backpropagation—now the workhorse of deep learning.*4 Backpropagation works by estimating th

Finally, because of the scale at which current AI can operate, there are many ways in which AI could (even in its still-primitive form) be used deliberately to cause serious public harm.