Researchers at Penn State have developed an electronic tongue capable of detecting minute variations in liquids, from milk to coffee and fruit juices. This cutting-edge technology not only enhances food safety and production but also provides a unique window into the decision-making process of artificial intelligence (AI). By mimicking the human gustatory cortex, the electronic tongue and its AI can classify and assess the quality, authenticity, and freshness of various substances with remarkable accuracy.
Finding the Mystery Behind Taste
Our tongues are amazing, sensitive things; we can detect far more than before imagined by the billions of flavor-sensitive taste buds embedded on our tongues. The tongue is, indeed, the most prominent sensory organ experiencing taste…. When food comes into contact with taste receptors, the data is sent off to a part of the brain that interprets these flavours, known as an olfactory bulb.
To replicate that biological procedure, the Penn State team created an electronic tongue made of a graphene-based ion-sensitive field-effect transistor—an electronic device capable of sensing ions in chemical solutions. Next, connect it to an artificial neural network that was previously trained on different datasets to characterise the features of the two liquids.
AI Meets the Taste Buds
The investigators gave the neural network 20 concrete variables to test — all concerning properties of how the liquid sample interacts with the electrical properties of the sensor. Using these human-chosen numbers, the AI was able to identify a range of liquids (such as different dilutions of milk, various sodas, coffee blends and fruit juices ranging from fresh to spoilt) with more than 80 per cent accuracy within a minute.
However, the researchers wished to take that even further. So, they chose to give the raw sensor data to the neural network and have it discover its own performance measures. Even more surprisingly, the accuracy of this neural network was almost 95% when using them instead of those supplied by researchers involved in their data collection. In doing so, this finding cracks open the black box that is the AI decision-making process, offering rare insight into how machines really learn from data.
Deciphering the Inner Thought Process of The AI
To interpret how the neural network made decisions, a technique called Shapley additive explanations was used. Utilizing this technique is a form of game theory (which in like way means thinking about what various players may do when choosing how they should carry on), in which data focuses are evaluated dependent upon the choice that is made by the evaluator.
From these explanations, the researchers could reverse engineer how much the neural network favored different aspects of measurements on your sample to make a final assessment. What they discovered is that the neural network was making its decision based not on a handful of single human-assigned parameters, but on the information it found important for any given position. This more expansive method enabled the AI to offset any day-to-day fluctuations and provide insight into changes like milk’s water content, or whether changes in indicators of spoilage were significant enough to pose a food safety risk.
