Introduction
In a lab situated near Newark Airport, Osmo, an innovative startup, is striving to endow computers with the ability to perceive and analyze scents. With a team of around 30 trained human sniffers, Osmo is documenting a vast range of odors, meticulously categorizing them with descriptive labels such as “nutty,” “peachy,” and “marine.” This endeavor aims to construct a foundational model for smell, which Osmo claims to be the first of its kind.
The Science of Smell
According to founder Alex Wiltschko, scent is a fundamental language shared by many species, enabling communication that transcends the limitations of visual and auditory interactions. The powerful association between smells and memories illustrates the profound impact that olfactory experiences can have on human emotion and cognition. Wiltschko’s ambition is to make computers capable of understanding this complex sensory information.
Achievements and Applications
Wiltschko asserts that Osmo’s AI model has reached a level of olfactory intelligence comparable to that of an average human. The model is proficient not only in detecting and distinguishing between various scents but also in formulating new fragrances, reportedly outperforming professional perfumers at a leading consumer goods company. Possible applications for this technology include enhancing threat detection systems in airports or aiding in medical diagnostics.
Data Challenges and Solutions
Osmo’s journey began with the challenge of insufficient data, as existing databases contained fewer than 200,000 human scent descriptions. To overcome this barrier, the company has created its own extensive dataset, comprising over 10 million human labels and approximately 800 million “scent tokens.” These tokens detail the specific molecules and their concentrations that contribute to distinct odors. Using advanced spectrometers and chemical sensors, Osmo continuously analyzes a diverse array of materials, from fruits and flowers to everyday pantry items.
The Role of Human Labelers
To ensure the accuracy and richness of its dataset, Osmo employs around 70 trained data labelers who engage in the meticulous process of smelling various substances and documenting their characteristics using a standardized vocabulary. This rigorous labeling process is crucial for the AI model’s ability to learn and generalize from the data.
Insights from Scent Composition
A fascinating insight derived from Osmo's research is the parallel between linguistic structure and scent composition. Just as the arrangement of words can alter the meaning of a sentence, the combination and proportion of scent molecules can significantly influence the perception of odors. For instance, Wiltschko notes the surprising similarity between the scent tokens associated with a newborn baby and those linked to a plate of cooked spaghetti, demonstrating the intricate relationships between different smells.
Innovative Scent Creation
Osmo has also developed a unique scent-printing robot that can generate fragrances based on specific customer requests. This robot operates similarly to a 3D printer, mixing ingredients according to exact formulas to create samples tailored to particular prompts. During a demonstration, Wiltschko showcased the AI’s capability to generate scent profiles based on user-defined criteria, further highlighting the potential for customized fragrance creation.
Collaborations and Grants
The technology developed by Osmo has attracted interest from major corporations, including Procter & Gamble, for whom it has created retail fragrances. Additionally, the Gates Foundation invested in the startup, providing $5 million initially and a subsequent grant of $3.5 million to explore molecules that could effectively repel disease-carrying insects. Osmo has also collaborated with StockX, utilizing scent analysis to help identify counterfeit sneakers by examining chemical discrepancies in materials.
Skepticism and Criticism
Despite its advancements, Osmo faces skepticism from some experts in the field of olfaction. Avery Gilbert, a smell scientist, has raised concerns regarding the validity of Osmo’s claims, suggesting that the reliance on human sensory panels for patent filings contradicts the premise of an AI-driven solution. Furthermore, Michael Schmuker, a former professor of neural computation, acknowledges the technological achievements of Osmo but warns that the complexity of human olfaction might not be fully captured by the current model. He emphasizes the significance of context in smell perception, noting that the same molecule can evoke different associations based on situational factors.
Conclusion
Osmo’s efforts to merge AI with olfactory science represent a significant step forward in the understanding and application of scent technology. While the company has made notable strides in creating an AI capable of mimicking human olfactory abilities, ongoing challenges related to data, perception, and contextual understanding highlight the complexity of this endeavor. As Osmo continues to develop its technology, the intersection of AI and olfactory science may lead to innovative applications across various industries, transforming how we experience and interact with the world of smells.
Source
technologyreview.com