Have you ever wondered whether that old pack of fruit juice in the back of your fridge is still safe to drink? A new ‘electronic tongue’ might just tell you.
The system leverages artificial intelligence (AI) to identify food safety and freshness issues. It also provides a glimpse into how AI makes decisions, researchers reported in the October 9 issue of Nature.
To create the tongue, the researchers used ion-sensitive field-effect transistors, devices that detect chemical ions. The sensor collects information about the ions in the liquid and converts that information into electrical signals that can be interpreted by a computer.
“We are trying to create an artificial tongue, but the tongue is not the only part of the process of how we experience different foods,” study co-author Saptashi Das, an engineer at Pennsylvania State University, said in a statement. It’s not related,” he said. “Our tongue itself is made up of taste receptors that interact with food species and send that information to the gustatory cortex (a biological neural network).”
Related: Robot hand looks very ‘human-like’ thanks to new 3D printing technology
In the new system, the sensor plays the role of the tongue, and the AI plays the role of the gustatory cortex, the brain region that recognizes taste. The researchers linked the sensors to an artificial neural network, a machine learning program that mimics the way the human brain processes information, to process and interpret the data collected by the sensors.
Initially, Das and his colleagues fed a neural network several parameters to use when determining how acidic a particular liquid is. Using these parameters, the neural network determined acidity with approximately 91% accuracy. Having the neural network define its own parameters for acidity analysis increased accuracy to over 95%.
Next, we tested our tongues with real drinks. The system distinguishes between similar soft drinks and coffee blends, assesses whether milk is diluted, identifies when fruit juice has gone bad, and detects harmful per- and polyfluoroalkyl substances in water. Researchers have discovered that they can detect PFAS.
Using an analysis technique called Shapley Additive Explains, the researchers were able to determine which parameters were most important in helping the neural network reach its conclusions. The researchers say the technique could help scientists understand how neural networks make decisions, but this remains an open question in AI research.
“We found that the network was looking at more subtle characteristics of the data, things that we as humans struggle to adequately define,” Das said in a statement. “Neural networks also take into account the sensor characteristics holistically, reducing the fluctuations that can occur from day to day.”
The ability to adjust for these variations could help make sensors more robust in other applications. Throughout the decision-making process, the neural network takes into account variations that currently reduce the reliability of ion-sensitive field-effect transistors in some situations.
“We realized that we can live with imperfections,” Das said in a statement. “And that’s what nature is. Nature is full of imperfections, but like our electronic tongues, it can make solid decisions.”
