Algorithms are Unable to Question Data

Humans have an amazing ability; We are able to reason about things that do not exist anywhere but in our minds. We exist simultaneously in a physical and an imagined reality. Our ability as a society to establish a shared understanding of reality is what has lead us to dominate the world. It enables us to use money, invent things and trust each other.

The most widely reported breakthroughs in machine learning in recent years have been in supervised learning. Given enough data state-of-the-art supervised learning algorithms can, without question, approximate behaviour very well.

However, the ability to replicate behaviour is not how I think of intelligence. Supervised learning algorithms currently lack the ability to establish a shared understanding. Inherent in establishing a shared understanding is questioning data and negotiating a mutually shared interpretation of events.

Don’t get me wrong, supervised machine learning is extremely useful. It will eliminate boring tasks and free up resources better spent elsewhere. But it will not lead to artificial general intelligence. Reading the news right now you might be confused and think artificial general intelligence is just around the corner. It is not.

Some might say “but what about AlphaGo?”. I think AlphaGo is an amazing scientific and engineering accomplishment. It makes me really excited. At the same time solving a perfect information game where the goal is very clear, i.e. score the most points according to very specific rules, is not the same as reasoning under uncertainty the way humans do in their daily lives.

The main bottleneck for supervised machine learning is that it requires you to specify the desired output through examples. Producing consistent examples describing the desired output is challenging as it requires negotiating interpretations of data, questioning what is correct.

Let this be an introduction to something I will be writing more about over the coming months. I am, along with a great team, in the process of building a platform to help people establish a shared understanding of data.

We believe that even in a world pervaded by intelligent machines, people will remain invaluable because of their judgment.