A few years ago, we ‘only’ had to worry about things like climate change and the unsustainable pressures on the healthcare system. Then suddenly, the threat of Artificial Intelligence (AI) joined the collection of things we need to worry about. Luminaries such as Elon Musk, Stephen Hawking and Bill Gates have expressed concerns that AI is a threat to our jobs, and possibly to our very existence.
How worried should we be about AI?
The term AI was coined in the mid-1950’s and has continued to trigger angst in our collective psyche ever since. Part of the problem is the name; “Artificial Intelligence” is a loaded term. We see intelligence as the primary difference between us and everything else, particularly machines, so artificial intelligence is threatening.
The fact that computers can perform better than humans in clinical medicine has been documented for over 40 years. However, these systems have been in narrow fields. For example, in 1972 de Dombal  demonstrated a computer program was better than senior clinicians at diagnosing patients with acute abdominal pain (91.8% vs 79.6% accuracy). This amazing result didn’t stir apocalyptic predictions for humanity, I suspect because the title of the paper was “Computer-aided Diagnosis of Acute Abdominal Pain” rather than “Artificial Intelligence is superior to doctors”.
So why has AI risen to prominence recently? And why are people worried about it?
Types of AI
AI systems can be broadly thought of as knowledge-based, and connectionist. Knowledge-based systems are the traditional approach to AI from the 1950’s. They require hand-coding facts and relationships about the world and then run rules to update models and generate new facts or decisions. These systems are laborious to build and maintain, they don’t handle uncertainty well, but they can work well in narrow areas of specialty.
The other broad method in AI is the connectionist approach that uses data to train models with minimal human intervention. This includes neural networks (NNs). While NNs have also been around since the 1950’s they came into prominence in the 1980’s when they worked very well for handwriting recognition. Interest in NNs fizzled out in the 1990’s when they couldn’t scale to more complex recognition tasks.
The Rise of Neural Networks and Deep Learning
Around 2012, something dramatic happened. After years in the wilderness NN based systems made huge leaps in image recognition and speech recognition within a single year, beating all other approaches. Prior to 2012 the best performance in image recognition had a 25% error rate, in 2012 this dropped to 15% (Figure 1). Since 2015 NNs can perform better than humans at recognising objects in pictures (image recognition).
Figure 1. The improvements in Image recognition error rates using deep neural networks.
Human performance is 5% error which was overtaken by neural networks in 2015.
Today’s NNs are very different from previous generations. While they still use the same basic methods of the past they are deep: they have hundreds of layers of ‘neurons’ they can use to learn from data. Training these networks is coined ‘deep learning’.
The rapid rise in performance of deep NNs (DNNs) is driving the most recent concerns about the potential impact of AI on jobs and, more generally, our demise as a species.
Three things came together to make DNNs the dominant form of AI today:
- Improvements in NN techniques that allowed networks to have many layers;
- Speeding up training by using processors originally used for video games, so DNNs could be trained in days instead of months;
- The availability of large volumes of data with which to train DNNs.
DNNs allow machines to do tasks that were otherwise beyond them, such as image recognition, speech recognition, and language translation. However, it’s important to note that DNNs are still narrow expert systems designed for specific tasks. Learning in one area, such as recognizing positions on a Go Board, does not translate to other fields like helping us diagnose patient X-rays.
Occasionally I get asked if deep learning will result in killer robots. My current answer is that DNNs will be the eyes and ears of the robots, but they will still require some other yet-to-be-developed technology to really plan an effective revolution. So that’s good.
In my next post, I’ll look at how DNNs and AI are performing in healthcare today.
- de Dombal FT, Leaper DJ, Staniland JR, McCann AP, Horrocks JC. Computer-aided diagnosis of acute abdominal pain. Br Med J. 1972;2(5804):9-13.
Download our free whitepaper to learn how AI can enhance your hospital data to better understand patient complexity.Download the whitepaper