December 4, 2019 | When you think of artificial intelligence, what comes to mind? A computer brain? Computer vision? Auditory language processing? What about smell?
It’s an area of research Luisa Bozano is particularly intrigued by. “It is a really fascinating AI problem, and while most of the models are vision-inspired neural network based, very few, even for the olfaction IoT platforms, are currently using an AI model based on how organisms process the chemical information that enable us to smell.”
Bozano is digging into this area at IBM, where she’s the Manager of Nanoscale Fabrication. On behalf of Bio-IT World, John Koon recently spoke to Bozano about IBM’s research into an AI-enabled nose, the applications of olfactory sensing, and the research behind it all.
Editor’s Note: Koon is a technology editor, writer, and researcher at Tech Idea Research. He’s chairing a session on IOT Platforms For Diagnostics And Remote Monitoring at Cambridge Innovation Institute’s Sensors Summit in San Diego next week. Work from Bozano’s group at IBM will be presented. Their conversation has been edited for length and clarity.
Bio-IT World: When the term machine learning or artificial intelligent (AI) is brought up, the first word that comes to mind is “Watson” from IBM. Are you part of the Watson team? How is your group organized?
Luisa Bozano: This project is coming out of the Almaden Research Center in San Jose, California, that is one of the 12 worldwide laboratories of IBM Research spread across six continents. The work that we do is exploratory research, and is not a part of our commercial technology portfolio.
Our research was driven by the desire for innovation, using existing expertise and unique IBM know-how to combine the broad range of skills of our teams: from materials science to engineering and data analytics.
With this work, we see an opportunity to implement an AI-enabled nose in the future, which could potentially lead to enabling AI platforms such as Watson with smell capabilities.
Sensors have broad applications in many fields including medical, automation, manufacturing, automotive, agriculture, and others. It seems that only recently olfaction has been mentioned. Can you share the history of why IBM wants to study olfaction?
Our research team has approached studying olfaction from a bottoms-up perspective. We knew if we wanted to create solid AI models, we first needed to obtain quality data. This is often an issue for AI-enabled IoT devices, as they need stable and reliable data.
With olfaction the standard model is to use a platform based on partially specific sensors. Each sensor is sensitive to multiple gases, and a single gas can be detected by multiple sensors. However, the overall electrical response of the sensors has to be different to be able to build a good “fingerprint,” or a unique signature for that smell.
Currently, there are many sensors based on the fingerprint model, however they all have many issues related to drift and stability. For this project, we leveraged IBM’s long history in materials science and hardware development to bring together new and innovative solutions around sensors. We are currently working on chemiresistors MOS (metal oxide semiconductors), as well as developing new organic compound materials.
It is a really fascinating AI problem, and while most of the models are vision-inspired neural network based, very few, even for the olfaction IoT platforms, are currently using an AI model based on how organisms process the chemical information that enable us to smell.
What are some of the applications of olfaction sensor IoT platforms?
The power of IoT platforms is due to their accessibility, portable nature, and cloud data access. However, the highest value for our platform is what it is capable of detecting.
Volatile organic compounds (VOCs) are everywhere around us. They are in the air we breathe, they are in our food, they are in our breath, they can be markers for our health and for the quality of the food we eat or sell. Plants release VOCs when sick, if attacked by insects to warn nearby plants, and also as a self-defense mechanism. They are the markers for the safety of the environment that surrounds us, alerting us if there are contaminants, irritants, or allergens.
We see the potential in the combination of this knowledge and information, along with its ease of the measurement (measuring smell is not invasive, and it is relatively fast compared to alternative techniques), to open doors for environmental, food, healthcare applications.
For food quality, we are very interested in using IoT platforms for blockchain applications. The real value is using the IoT platform as a crypto anchor or digital fingerprint that can prevent fraud and tampering of a product.
Can you briefly describe what IoT platforms for olfaction are?
The term olfaction applies to chemisensory systems that detect volatile molecules generated at a distance. These molecules of interest are often found in low concentrations in the environment (e.g., parts per million or parts per billion) and come mixed-in with a variety of other molecules, whose presence may confound identification.
A widely-investigated strategy for odor identification relies on utilizing an array of sensors, in which each device responds in complex ways to the presence of gases, and differently from the other devices comprising the array. Within the array, each sensor may respond to various molecules, and each molecule can be detected by multiple sensors. The collective output of the array is then processed to generate a “fingerprint”, or a unique set of responses that can be associated by similarity (and the help of machine learning algorithms) to a known pattern, and therefore to a given odor.
IoT platforms for olfaction often follow this approach to recognition, while offering the additional range of benefits associated with the internet connectivity and cloud access, as well as the advantages of portable platforms.
What is involved in applying machine learning to the olfactory platforms?
When it comes to olfactory IoT platforms, data preprocessing and pattern recognition techniques go hand-in-hand when successfully differentiating one odor from the next.
The main component is understanding the data and your sensors, so the first step is usually data visualization and exploration, using methods like Principal Component Analysis (PCA) and t-distributed Stochastical Neighbor Embedding (t-SNE).
Depending on the application and your sensors, such as a high spread in your data or drift in your sensors, the analytical approach may be adapted. A variety of approaches exist for electronic noses, but we have developed our proprietary methods to work with our platform that highlights the important information from our sensors. Once the data has been preprocessed, the application of pattern recognition and machine learning also varies, with the use of simple classifiers like support vector machine and logistic regression, or vision-based artificial neural networks. There are pros and cons for each approach, but it depends on the experimental setup and the intended application, so it is an art which requires experimentation with these methods to build intuition.