We rely on experts at all times. If you need financial advice, you ask an expert. If you are sick, you go to the doctor, and as a jury member you can listen to an expert witness. However, in the future, artificial intelligence (AI) may replace many of these.
In forensic science, expert witness plays an important role. Lawyers seek them out for their analysis and opinion on expert evidence. But experts are human beings, with all their failings, and the role of expert witnesses is often associated with a miscarriage of justice.
We are investigating the potential of AI to study evidence in forensic science. In two recent papers, we found that AI was better at assessing footprints than general forensic scientists, but not better than typical footprint experts.
What’s in the Footprint?
You leave footprints when you walk around your house barefoot, such as on your carpet or in the form of residue from your feet. Bloody footprints are common at violent crime sites. They allow investigators to reconstruct events and perhaps profile an unknown suspect.
Shoe prints are one of the most common types of evidence, especially in home theft. These marks have been recovered from windows, doors, toilet seats and floors and may be visible or hidden with the naked eye. In the UK, the points recovered are analyzed by police forces and used to search databases of shoe patterns.
The size of a barefoot print can tell you about a suspect’s height, weight, and even gender. In a recent study, we asked an expert podiatrist to determine the gender of a group of footprints and they corrected it just over 50% of the time. We then built a neural network, a form of AI, and asked it to do the same thing. Got it right about 90% of the time. What’s more, to our surprise, it can also assign an age to the track-maker, at least to the nearest decade.
When it comes to shoe prints, shoe experts can only identify the make and model of the shoe from experience – it’s second nature for these experts and mistakes are rare. Anecdotally, we are told that there are fewer than 30 footwear specialists in the UK today. However, there are thousands of forensic and police personnel in the UK who are casual users of footwear databases. For these casual users, analyzing shoes can be challenging and their work often needs to be verified by an expert. For this reason, we thought AI could help.
We commissioned a second neural network, developed as part of an ongoing partnership with UK-based Bluestar Software, to identify the make and model of footwear impressions. This AI takes the impression of a black and white shoe and automatically recognizes the size of the component threads. Are the components square, triangular or circular? Is there a logo or writing on the imprint of the shoe? Each of these figures corresponds to a code in a simple classification. It is these codes that are used to search the database. In fact the AI gives the user a series of suggested codes to verify and identify areas of ambiguity that need to be checked.
In one of our experiments, an occasional user was given 100 randomly selected shoe prints to analyse. During the test, which we ran several times, the casual user got it correct between 22% and 83% of the time. In comparison, AI was successful between 60% and 91%. However, footwear experts are almost 100% correct.
Our second neural network was not perfect and one of the reasons it did not outperform real experts is that shoes change with wear, making the task more complex. Buy a new pair of shoes and walking is fast and clear but after a month or two it becomes less obvious. But while AI couldn’t replace an expert trained to find these things, it occasionally outperformed users, suggesting that it could help the expert spare time to focus on more difficult matters. .
Will AI replace experts?
Such systems increase the accuracy of footwear evidence and we will probably see it used more often than currently – especially in intelligence-led policing that aims to link crimes and reduce the cost of home burglary. In the UK alone they cost an average of £5,930 per incident in 2018, the equivalent of a total economic cost of £4.1 billion.
Read more: It takes a lot of energy for machines to learn – here’s why AI is so power-hungry
AI will never replace the skilled and experienced judgment of a trained footwear tester. But it can help ease the burden on those specialists and allow them to focus on difficult cases by helping casual users to more reliably identify the make and model of a footprint on their own. At the same time, experts using this AI will replace those who do not.