# Limitations in odour recognition and generalisation in a neuromorphic olfactory circuit

**Authors:** Nik Dennler, André van Schaik, Michael Schmuker  
**arXiv:** [2309.11555](https://arxiv.org/abs/2309.11555)  
**Date:** 2023-09-20  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2309.11555v1/

## Key points

- Imam & Cleland developed an odour-learning algorithm based on neuromorphic computing principles inspired by mammalian olfactory bulb neural pathways
- Algorithm utilized Metal Oxide (MOx) gas sensors and spike-time dependent plasticity rules for creating gas representations
- Model validated for learning and identifying gases using Jaccard similarity coefficient, but limitations were identified
- Major limitation was the dataset used for validation, which had sensor drift and non-randomized measurement protocol issues
- Model demonstrated robust recognition of certain gases like Toluene but had limitations in generalizing over repeated presentations of the same gas
- A simple hash table approach matched or exceeded reported results in accuracy and runtime for odour identification tasks
- Further validation beyond restoring learned data samples is necessary to assess suitability for real-world applications involving unknown contaminants and sources of interference

## Summary

In their study, Imam & Cleland developed an odour-learning algorithm based on neuromorphic computing principles inspired by the neural pathways of the mammalian olfactory bulb. The algorithm utilized recordings from Metal Oxide (MOx) gas sensors to create representations of gases through spike-time dependent plasticity rules applied to sequential gamma-frequency spike packages. The model was validated for its ability to learn and identify gases by comparing clean gas representations with artificially occluded sensor recordings using the Jaccard similarity coefficient. However, limitations were identified in the study that raised concerns about the validity of its conclusions. One major limitation was the dataset used for validation, which suffered from sensor drift and a non-randomized measurement protocol. The lack of randomization in gas presentations over time during recording, along with the presence of sensor drift contaminations, rendered the dataset unsuitable for accurate classification tasks. In replicating parts of the study, it was found that while the model demonstrated robust recognition of certain gases like Toluene, it had limitations in generalizing over repeated presentations of the same gas. Additionally, a simple hash table approach was able to match or exceed the reported results in accuracy and runtime for odour identification tasks. Overall, while Imam & Cleland's algorithm showed promise in utilizing neuromorphic computing for odour learning and identification tasks, further validation beyond restoring learned data samples is necessary to assess its suitability for real-world applications involving unknown contaminants and sources of interference. Additional considerations such as addressing sensor drift issues and improving generalization capabilities are crucial for advancing the effectiveness of neuromorphic computing in Machine Learning and Artificial Intelligence.