Neuromorphic olfaction systems are bio-inspired computational architectures that mimic the neural processing mechanisms of biological olfactory systems using neuromorphic engineering principles. These systems combine chemical sensors with spiking neural networks (SNNs) to process and classify odors in an event-driven, power-efficient manner similar to how biological brains process smell. [1]
The development of neuromorphic olfaction systems emerged from the convergence of electronic nose technology and neuromorphic computing in the early 2000s. The NEURO-CHEM project (2008–2011), funded by the European Union's FP7 program (Grant Agreement Number 216916), brought together researchers from multiple institutions to develop novel computing paradigms and biomimetic artifacts for chemical sensing. [2] This project resulted in significant advances including the development of a large-scale chemical sensor array with 65,536 sensor elements. [3]
Neuromorphic olfaction systems operate by implementing computational models inspired by the mammalian olfactory bulb. The architecture includes Olfactory Receptor Neurons (ORNs) that convert chemical signals into electrical spikes, Projection Neurons that process and relay olfactory information, Lateral inhibitory neurons that provide pattern separation and noise reduction, and Spike-Time Dependent Plasticity (STDP) that enable online learning capabilities. [4]
The systems use event-based processing, where sensory data is represented as sparse spikes that encode critical information for classification and identification of odors. [4]
Several neuromorphic hardware platforms have been used to implement olfaction systems:
In 2020, Intel and Cornell University researchers demonstrated a neuromorphic olfactory circuit on the Intel Loihi chip that could learn and recognize scents of 10 hazardous chemicals. The implementation utilized distributed, event-driven computations and achieved rapid one-shot online learning. [5] Intel's team configured "a circuit diagram of biological olfaction" on Loihi using a dataset of activity from 72 chemical sensors. [6]
BrainChip's Akida neuromorphic processor has been used for rapid classification of multivariate olfactory data. The system demonstrated dynamic power consumption of only 24.5 mW with high throughput of 181 inferences per second when applied to malt classification. [7]
Neuromorphic olfaction systems have potential applications across multiple domains:
The systems can detect biomarkers in breath for disease diagnosis, including detection of volatile organic compounds associated with various medical conditions.
Applications include detection of pollutants, hazardous gases, and monitoring of air quality in industrial and urban environments.
Quality control applications include classification of malts for brewing, detection of food spoilage, and authentication of food products. [8]
Active research areas in neuromorphic olfaction include:
Key challenges facing neuromorphic olfaction systems include:
