Can We Trust Artificial Intelligence?

Deep learning is used in everything from speech recognition software to the assessment of mortgage applications. The only trouble is, we don’t really know how it works…

Deep learning does it all: face recognition, speech recognition, language translation, automated game playing. It’s an approach that has transformed the field of artificial intelligence (AI) and is without doubt the AI flavour of the decade. It works amazingly well.

But increasingly, questions are being asked about transparency. Exactly how does deep learning work? Can we trust it to work for safety critical applications such as self-driving cars? We generally like computer algorithms to be as transparent and revealing as possible – and this is not the case when it comes to deep learning.

In essence, deep learning is really a clever rebranding of an earlier computer learning method called artificial neural networks (ANNs). Dating back to the beginning of computers, ANNs are computer programs that simulate networks of neurons like those in our own brains. They’re hugely simplified and don’t really work in quite the same way that real neurons work, but nevertheless, they enable computers to learn (see ‘How a neural network works’, opposite).

HIDDEN DEPTHS

Research on neural networks started in the 1950s and the ideas were refined over many decades, but it became clear that neural networks were not as good as some other approaches in machine learning (the branch of AI dedicated to helping computers learn from data in order to make classifications and predictions). As a result, research in the area began to wane by the early 1990s, and learning methods that relied on clever statistics started to dominate.