Teaching AI to Smell
How Cogito Tech Enabled Fragrance Recognition Through Expert-Labeled Olfactory Data
Challenge
A global leader in the fragrance and sensory AI space approached Cogito Tech with a unique challenge. They wanted to train their AI model to accurately recognize, classify, and evaluate fragrances as perceived by humans. Unlike visual or audio data, olfactory data is inherently subjective, making consensus-building for model training extremely complex.
Solution
Our team implemented a structured multi-dimensional olfactory evaluation framework involving intensity & pleasantness scaling, grand family identification and ranking, and comparative experiments (Dissimilarity & XAB) to generate high-quality, consistent scent perception data, which made digital olfaction a viable and scalable reality.
Outcome
We successfully completed over 200,000 hours of odor configuration and annotation, following rigorous protocols designed for accuracy and consistency. By complying with these standards, Cogito Tech achieved 100% accuracy in annotations, determining reliable and effective training data for the client’s AI model.
How We Made It Work?
Step 1 – Building a Standardized Fragrance Database
We created a structured dataset reflecting human perceptions across intensity, pleasantness, and fragrance families.
Step 2 – Overcoming Annotation Challenges
Unlike images or sounds, smells can’t be replayed. Fragrance perception varies greatly across individuals and cultures, making annotation highly complex.
Step 3 – Implementing a Robust Methodology
- Olfactory Annotator Recruitment: We built a trained panel of expert sniffers.
- Controlled Lab Environment: Rigorous scent evaluation under standardized temperature and air conditions.
- Consensus-Driven Annotations: Cultural, age, and gender diversity were factored into pleasantness scoring to minimize bias.
Cogito Tech’s Solution
To overcome the challenge, we strategize to work on following core attributes:
1. Intensity Scale (0 to 10)
- Annotators rated the strength of the scent on a scale from 0 (barely perceptible) to 10 (extremely intense).
- It helped build a quantifiable metric for olfactory signal strength.
2. Pleasantness Score (0 to 10)
- Annotators sniffed how pleasant or unpleasant a scent was based on personal perception but normalized across diverse demographics.
- Cultural sensitivity and gender/age diversity were considered to avoid bias.
3. Fragrance Taxonomy (Family Classification)
- Our team tests custom taxonomy of 11 fragrances families. The list includes, woody, animalic, citrus, floral, fruity, green, herbal, industrial, mineral, soulful, and sweet/balsamic.
Results Delivered
Over 200,000 hours of odor configuration and annotation completed with 100% accuracy, enabling consistent model training.
Client’s AI fragrance engine became capable of detecting and rating fragrance intensity. It started classifying smells into predefined families, and predicting end-user preferences regarding fragrance based on model output.
Do you have complex, subjective data you need to make machine-readable? Let’s connect.