As you say the use cases are virtually limitless in the olfactory space. Brainchip has been proving out AKIDA’s ability in this field since at least 2019. The following make for interesting reading:
5. Conclusions
In this paper, we presented the development of an SNN-based solution for real-time classification of electronic olfactory data. The SNN was implemented using Brainchip’s Akida Development Environment, which provides an emulation of the Akida NSoC’s functionalities on a Python-based
Sensors 2019, 19, 4831 11 of 13
software platform. The highlights of this implementation included the development of a novel AER-based encoder for olfaction data, implementation of unsupervised STDP for training the SNN, highly accurate real-time classification results, and preliminary results that lay a foundation for applying the Akida SNN for reliable early classification results.
One of the most significant contributions of this study is the development of AERO, an AER-based data-to-spike encoder for olfaction data. The operating principle of AERO is based on discretizing the sensor responses and encoding their activation levels along with sensor ID and temporal data. For this implementation, normalized relative resistance features were extracted from sensor responses and provided as an input to the AERO encoder. Parameters, such as frequency of event-generation, the number of time points for encoding, and discretization levels, can be configured based on the input data and processing requirements. The development of AERO has opened several avenues for future research, such as encoding multi-dimensional data using different features and interfacing of electronic nose systems with AER-based neuromorphic hardware for processing.
The SNN was tested using the benchmark dataset [20] under four different scenarios of increasing complexity. In general, under each scenario, the classification performance of the SNN was between 90% and 100%, and the processing latency was between 2.5 and 3 s, which includes data-to-spike encoding, learning, classification, and other software-based latencies introduced due to looping and conditional statements. This processing latency would, of course, be dramatically reduced once the classifier is implemented on the Akida NSoC hardware without the overhead of software emulation. Taken together, these classification results show that the SNN-based classifier can deliver highly accurate results with minimal processing latency. Moreover, the ability to transfer the SNN implementation to the Akida NSoC can be leveraged to develop low-power electronic nose systems with minimal computational cost and memory requirements. Intrinsically, in most cases, neuromorphic approaches have proven to outperform traditional processing methods that suffer from limited accuracy, high computational and power requirements, and substantial latency to provide classification results [32]. When evaluated against other neuromorphic and traditional approaches based on the same dataset [21,24,33,34], the results revealed that the SNN classifier developed in this study achieved comparable and, in most cases, better classification performance with minimal computation requirements and latency for both learning and processing. More importantly, the Akida SNN was able to identify patterns from a highly multi-dimensional dataset and classify the dataset based on the four chemical groups of the compounds.
Future research based on these results will focus on the development of a robust SNN-based classifier on the Akida NSoC and its implementation in a real-world application. The efficacy of AERO when combined with rate coding methods, such as [35] and rank-order encoding [2,31,36], will also be investigated. This implementation also lays the foundation for the application of both the AERO encoder and an SNN to study neuromorphic gustation and the fusion of olfaction and gustation for the development of a comprehensive analytical tool for chemical sensing
AND
“4. Conclusions
This study presents the implementation of a neuromorphic approach towards the encoding and classification of electronic nose data. The proposed approach was used to identify eight classes of malts and has potential as an application for quality control in the brewing industry. Experiments were conducted using a commercial e-nose system to record a dataset consisting of time-varying information of sensor responses when exposed to different malts under semi-laboratory conditions. The classifier proposed in this study utilized the combination of the Akida SNN and the AERO encoder, a neuromorphic approach that has previously delivered highly accurate results on a benchmark machine olfaction dataset [12]. The proposed method successfully classified the dataset with an accuracy of 97.08% and a maximum processing latency of 0.4 ms per inference when deployed on the Akida neuromorphic hardware. A secondary dataset that was used to validate the classifier model in an ‘inference-only’ mode was classified with an accuracy of 91.66%. These results could potentially be further improved by refinements to pre-processing that can enhance informative independent components for malt classes that are misclassified.
Based on these results, we can conclude that the classifier model implemented using Akida SNN in conjunction with the AERO encoder provides a promising platform for odor recognition systems. An application targeted towards the identification of malts based on their aroma profile, generally considered a nontrivial classification task using traditional machine learning algorithms, was successfully demonstrated in this work with a classification accuracy greater than 90% under different scenarios. The developed model can be deployed on the Akida NsoC, thus enabling the integration of a bio-inspired classifier model within a commercial e-nose system. A comparative analysis of the proposed approach with statistical machine learning classifiers shows that the SNN-based classifier outperforms the statistical algorithms by a significant margin for both accuracy and processing latency. A performance-based comparison of the neuromorphic model proposed in this work with other neuromorphic olfactory approaches, such as [13,14,26,27,69,70], could not be established as their inherent structures, including spike encoding schemes, neuron models, SNN architectures, and implementation of learning algorithms, vary vastly. The proposed methodology, however, does not require a graphic processing unit (GPU)-based model simulation, unlike in [13], or a complex bio-realistic model, as used in [14]. Furthermore, the SNN-based classifier can be entirely mapped on a single neural processing unit core, as opposed to multiple cores used in [14], leading to a low-power and low-latency implementation.
The application of such real-time and highly accurate e-nose systems can be extended to fields such as food technology, the brewing and wine industries, and biosecurity. Future research in this domain will focus on encoding parameters such as rank-order code within the AERO events to analyze its impact on classification performance.”
My opinion only DYOR
Fact Finder
(20min delay)
