Reflections on the nature and future of artificial intelligence

We organized the inaugural EY Analytics Summit at our Amsterdam office on Tuesday September 24th. During the event I shared some reflections on the nature and future of artificial intelligence (AI). Some of those reflections I’m also sharing through this blog as an attempt perhaps to go beyond the hype.

Is AI new?

The answer to this question is no. AI is not new and in fact has been around since the 1950s. What is new though in recent years, are the massive increases in processing power and storage capacity which have opened new doors for essentially old techniques. Many of the significant recent AI advances have been powered by what is referred to as deep learning, a subset of machine learning which allows systems to learn about the world from the data they absorb. Deep learning uses so-called artificial neural networks which are loosely based on our current understanding of how the biological brain works.

Supervised deep learning has led to significant improvements in areas such as computer vision, speech recognition and language translation. Supervised here means that the system learns a mapping, for instance between a set of pictures and a set of known labels (for example pictures of cats which are also labeled as cats). In essence, supervised deep learning is a statistical technique for finding correlations and classifying patterns in data. This also means that the models don’t have any real understanding of what they learn.

Does the brain really work like that?

Artificial neural networks are said to be loosely modeled after how the biological brain works. But does the biological brain really work like that though? Do humans really learn like that and require this many data examples? In the currently popular supervised deep learning approaches, systems need to learn from thousands and thousands, or even millions of examples, and all those examples need to be clearly labeled as well.

Humans though, do not need to study thousands or millions of examples to for instance identify a cat when they see one. Just one or a few examples will usually be enough. Also, past experiences teach us a lot about the world in general, which makes it easier to deal with new situations. As humans we are able to generalize, and transfer and apply our knowledge from one domain or situation to another which facilitates learning.

Do we know where AI is going to?

Supervised deep learning, can be seen as a road to achieving artificial intelligence. In recent years it is a well-traveled road which has led to many exciting developments. It is not the only road though and exploring other roads less traveled by, can open new horizons. Less known forms of machine learning, include transfer learning which fine-tunes existing models for new purposes and one-shot learning which requires just one or very few data examples. Combining machine learning with other forms of AI, can prove to be very valuable by providing the best of both worlds. Hybrid models also incorporating encoded logical rules and representations of the world, have the potential to be adaptive, while requiring less data and being more explainable than traditional implementations of neural networks.

We might only be able to 'see a short distance ahead', to quote Alan Turing, but in that vision of the future I do see a world of opportunity. AI certainly is a technology that expands what is possible for humans to do and it has the potential to spread very rapidly, as long as we are realistic about its current capabilities and the road that lies ahead.