Reproduce “SmellNet: A Large-scale Dataset for Real-world Smell Recognition” on OpenResearch

About this paper

Dewei Feng, Wei Dai, Carol Li, Alistair Pernigo, Yunge Wen, Paul Pu Liang

Researchers at MIT Media Lab and MIT EECS created SMELLNET, a large-scale dataset of real-world temporal smell signals collected from low-cost, portable sensors, paired with high-resolution chemical metadata. They developed SCENTFORMER, a Transformer-based model, which achieves 63.3% Top-1 accuracy for single substance classification and 50.2% [email protected] for seen mixture prediction, advancing machine olfaction with deployable sensor data.

Key insights

• Temporal differencing applied to sensor signals, emphasizing relative changes, significantly improves smell recognition accuracy by effectively capturing transient chemical dynamics.
• Models explicitly leveraging temporal structure, such as SCENTFORMER, CNNs, and LSTMs, consistently outperform non-temporal baselines, underscoring the importance of time-series analysis for sensor-based olfaction.
• Cross-modal training with GC-MS data can enhance sensor-only classification, particularly for less robust models, by providing complementary molecular structural information and improving model grounding.

Summary by alphaXiv.