In recent years, we have seen great strides being made in training artificial intelligence (AI) algorithms to understand language. These same algorithms have also proved themselves to be very useful in improving the performance of AI systems that process audio and visual inputs, which were already becoming quite advanced in their own right. When taken together, these technological advancements are giving machines the ability to understand the world in deeper ways than ever before.
The same level of effort given to understanding vision and language has not gone into processing other data modalities, such as the sense of smell, however. It may seem relatively insignificant by comparison — we have made it all these years without a smell-o-vision, after all — but scent is actually highly important in fully understanding the world around us. A machine with a keen sense of smell could sniff out the presence of allergens, diagnose diseases, and identify manufacturing processes that have gone awry before a complete failure, for instance.
A group of researchers at MIT has taken the first steps toward a future in which sniffing becomes a first-class artificial sense. They have developed SmellNet , which is a large-scale database containing digitized representations of a diverse range of smells in the natural world. Furthermore, the team used this dataset to train AI models to classify substances based on their smell (as captured by portable gas and chemical sensors) alone.
The SmellNet dataset includes 50 distinct substances, categorized across five types: nuts, spices, herbs, fruits, and vegetables. For each of these items, the researchers performed six 10-minute sensor readings using a set of carefully selected gas sensors. These sessions were spaced out over different days and controlled for ambient variables, resulting in approximately one hour of smell data per substance and 50 hours of data in total.
To capture the complex olfactory signatures, the team deployed a range of sensors. These included the WSP2110 semiconductor sensor for volatile organic compounds like acetone and benzene, the BME680 for temperature and humidity tracking, and the MQ-series and MP503 sensors for detecting gases such as ethanol, carbon monoxide, and ammonia. The Grove Multichannel Gas Sensor V2, with its independent metal-oxide channels, further broadened the spectrum of detectable compounds. An Adafruit ESP32 Feather board was used for sensor readout.
Using their dataset, the researchers trained a range of AI models for smell classification. The best-performing models included networks such as LSTMs and Transformers, and they incorporated preprocessing techniques like First-Order Temporal Differences to identify meaningful changes in gas concentrations over time. Additionally, the team employed contrastive learning methods to combine low-resolution sensor data with GC-MS chemical features.
The top model achieved up to 65.35% accuracy when classifying smells from pre-recorded sessions. When deployed in the real-world, under uncontrolled conditions, the results were much more modest. Accuracy scores of 10.71% to 25.38% were seen in these experiments.
Clearly there is still a lot of work yet to be done in this area, but the release of SmellNet could be the beginning of a new era of multisensory computing, where smell joins sight, sound, and language as a meaningful input for intelligent machines.The hardware captured rich information describing scents (📷: D. Feng et al.)
An overview of the models used in the study (📷: D. Feng et al.)
Source link
