Final Report Summary - BIOMACHINELEARNING (Bio-inspired Machine Learning for Chemical Sensing)
The BIOMACHINELEARNING project had three aims:
1) improving the speed and accuracy of odour detection and identification in electronic nose systems,
2) providing a bio-inspired solution for odour detection and classification using a bio-inspired network,
3) to implement this approach on a neuromorphic hardware system.
In order to improve the speed and accuracy of odour detection and identification in electronic nose systems, we first explored the temporal dynamics of gas sensor responses. To this end we analysed a publicly available large data set that contained responses of 6 models of gas sensors that were exposed to a turbulent gas plume inside a wind tunnel (Vergara et al., 2013). Sensors were placed at various locations down- and cross-wind from the gas source. We applied a bio-inspired band-pass filter approach that emphasizes temporal fluctuations in a wide frequency band, but suppresses electronic noise and slow signal fluctuations that occur due to switching the gas delivery on and off.
The resulting signal was processed to identify periods where it is monotonically rising, with the hypothesis that these “bouts” correspond to filaments in the gas plume hitting the sensor. We quantified the number of bouts depending on the relative location of the gas sensor to the gas source. Surprisingly, we found that the number of bouts gave a clear indication of the position of the sensor – The greater the distance to the gas source, the fewer bouts were encountered. Moreover, we found that the variance of bout counts during a time interval increased as a sensor was closer to the edge of the gas plume.
Our findings indicate that the information contained in the temporal dynamics of gas plumes can be exploited using off-the-shelf metal oxide sensors. This was previously not known and is a major result of this Fellowship. We have shown that these temporal dynamics enable an estimation of crosswind and downwind position of a sensor relative to a gas source in a turbulent environment. These results were published in Sensors and Actuators B: Chemical (Schmuker et al., 2016).
As a result of the successful dissemination of these findings in print, at conferences and lab visits, we are now working together with e-nose experts and a leading robotics group to test the utility of a method for gas-based robot navigation based on temporal fluctuations of gas concentration. If the follow-on work is successful, the findings obtained in this fellowship have potentially great impact on the use of robots in scenarios when visual navigation is not useful on its own, e.g. in disaster management. Other potential use cases for our research may lie in environmental monitoring and locating pertinent gas sources such as polluters or wild fires. Moreover, we have established collaborative work with a leading vertebrate olfaction lab with whom we investigate the neuroscientific aspects of decoding fast odour fluctuations in the olfactory system.
Towards building a bio-inspired network for odour detection and classification, we designed and implemented a neuromorphic network for odour recognition in electronic gas sensors. This was built on previous work (Schmuker et al., PNAS 2014) and was achieved in close collaboration with members of Prof. Nowotny’s group. The network achieved odour identification at an accuracy similar to state-of-the-art machine learning approaches (i.e. Support Vector Machines), and it runs on accelerated graphics cards (GPUs), via the GeNN simulator (Yavuz et al., 2016). We could show that a major advantage of the spiking approach to data classification is that it natively operates in real-time, facilitating online operation in parallel with data acquisition. The results have been published in Biomimetics & Bioinspiration (Diamond et al., 2016).
Towards the third aim of the project we implemented this approach on the SpiNNaker neuromorphic hardware system, this time not only aimed at odours and e-nose signals, but aimed at generic pattern recognition. We used the well-known MNIST handwritten digit recognition test and applied it on three spiking/neuromorphic platforms: 1) a massively parallel GPU simulation using the GeNN simulator on CUDA, 2) the “Spikey” system, a demonstrator board using the BrainScaleS hardware and 3) the SpiNNaker platform.
We found that all platforms achieved similar pattern recognition performance, in spite of the large differences in the underlying neuromorphic technology. The main focus of this study however was not so much on trying to achieve the best possible recognition performance (which has been achieved by conventional algorithms several years ago), but rather to assess the operation of hardware systems in a practical usage scenario. We could show that input/output bandwidth is a crucial limiting factor for the execution speed on current neuromorphic platforms. This results suggest to direct more effort towards the development of fast communication interfaces for these platforms. For the GPU and SpiNNaker platforms, we also analysed the power efficiency. A major result was that the host machine that handles data processing for these platforms consumed a significant portion of the total power. This implies that for best performance, the amount of computation and data processing that is carried out on the host machine should be minimised, and as many operations as possible should be offloaded to the devices. Such an approach will also minimise the need for host-device communication and attenuate the impact of bandwidth limitations. These results have been published in Frontiers in Neuroscience (Diamond et al., 2016).
Ongoing work based on these results consists in enabling real-time processing of e-nose signals with the SpiNNaker neuromorphic hardware system. To this end we have designed and built a low-cost gas sensor system that is based on the Arduino microcontroller platform. This system hosts 6 different metal oxide sensors. It can provide readings at intervals as short as tens of milliseconds. Our system enables us to perform our own e-nose experiments, and it serves as a platform to build a real-time interface to the SpiNNaker neuromorphic system and the pattern recognition system that we have implemented thereon.
Publications
Diamond, A, Schmuker, M, Berna, A Z, Trowell, S and Nowotny, Thomas (2016) Classifying continuous, real-time e-nose sensor data using a bio-inspired spiking network modelled on the insect olfactory system. Bioinspiration & Biomimetics, 11(2): 026002.
Schmuker, Michael, Bahr, Viktor and Huerta, Ramón (2016) Exploiting plume structure to decode gas source distance using metal-oxide gas sensors. Sensors and Actuators B: Chemical, 235: 636-646.
Diamond, Alan, Nowotny, Thomas and Schmuker, Michael (2016) Comparing neuromorphic solutions in action: implementing a bio-inspired solution to a benchmark classification task on three parallel-computing platforms. Frontiers in Neuroscience, 9: 491.
Other references
Vergara, A., Fonollosa, J., Mahiques, J., Trincavelli, M., Rulkov, N., and Huerta, R. (2013). On the performance of gas sensor arrays in open sampling systems using Inhibitory Support Vector Machines. Sensors Actuators B Chem. 185, 462–477.
Yavuz, Esin, Turner, James and Nowotny, Thomas (2016) GeNN: a code generation framework for accelerated brain simulations. Scientific Reports, 6: 18854.
1) improving the speed and accuracy of odour detection and identification in electronic nose systems,
2) providing a bio-inspired solution for odour detection and classification using a bio-inspired network,
3) to implement this approach on a neuromorphic hardware system.
In order to improve the speed and accuracy of odour detection and identification in electronic nose systems, we first explored the temporal dynamics of gas sensor responses. To this end we analysed a publicly available large data set that contained responses of 6 models of gas sensors that were exposed to a turbulent gas plume inside a wind tunnel (Vergara et al., 2013). Sensors were placed at various locations down- and cross-wind from the gas source. We applied a bio-inspired band-pass filter approach that emphasizes temporal fluctuations in a wide frequency band, but suppresses electronic noise and slow signal fluctuations that occur due to switching the gas delivery on and off.
The resulting signal was processed to identify periods where it is monotonically rising, with the hypothesis that these “bouts” correspond to filaments in the gas plume hitting the sensor. We quantified the number of bouts depending on the relative location of the gas sensor to the gas source. Surprisingly, we found that the number of bouts gave a clear indication of the position of the sensor – The greater the distance to the gas source, the fewer bouts were encountered. Moreover, we found that the variance of bout counts during a time interval increased as a sensor was closer to the edge of the gas plume.
Our findings indicate that the information contained in the temporal dynamics of gas plumes can be exploited using off-the-shelf metal oxide sensors. This was previously not known and is a major result of this Fellowship. We have shown that these temporal dynamics enable an estimation of crosswind and downwind position of a sensor relative to a gas source in a turbulent environment. These results were published in Sensors and Actuators B: Chemical (Schmuker et al., 2016).
As a result of the successful dissemination of these findings in print, at conferences and lab visits, we are now working together with e-nose experts and a leading robotics group to test the utility of a method for gas-based robot navigation based on temporal fluctuations of gas concentration. If the follow-on work is successful, the findings obtained in this fellowship have potentially great impact on the use of robots in scenarios when visual navigation is not useful on its own, e.g. in disaster management. Other potential use cases for our research may lie in environmental monitoring and locating pertinent gas sources such as polluters or wild fires. Moreover, we have established collaborative work with a leading vertebrate olfaction lab with whom we investigate the neuroscientific aspects of decoding fast odour fluctuations in the olfactory system.
Towards building a bio-inspired network for odour detection and classification, we designed and implemented a neuromorphic network for odour recognition in electronic gas sensors. This was built on previous work (Schmuker et al., PNAS 2014) and was achieved in close collaboration with members of Prof. Nowotny’s group. The network achieved odour identification at an accuracy similar to state-of-the-art machine learning approaches (i.e. Support Vector Machines), and it runs on accelerated graphics cards (GPUs), via the GeNN simulator (Yavuz et al., 2016). We could show that a major advantage of the spiking approach to data classification is that it natively operates in real-time, facilitating online operation in parallel with data acquisition. The results have been published in Biomimetics & Bioinspiration (Diamond et al., 2016).
Towards the third aim of the project we implemented this approach on the SpiNNaker neuromorphic hardware system, this time not only aimed at odours and e-nose signals, but aimed at generic pattern recognition. We used the well-known MNIST handwritten digit recognition test and applied it on three spiking/neuromorphic platforms: 1) a massively parallel GPU simulation using the GeNN simulator on CUDA, 2) the “Spikey” system, a demonstrator board using the BrainScaleS hardware and 3) the SpiNNaker platform.
We found that all platforms achieved similar pattern recognition performance, in spite of the large differences in the underlying neuromorphic technology. The main focus of this study however was not so much on trying to achieve the best possible recognition performance (which has been achieved by conventional algorithms several years ago), but rather to assess the operation of hardware systems in a practical usage scenario. We could show that input/output bandwidth is a crucial limiting factor for the execution speed on current neuromorphic platforms. This results suggest to direct more effort towards the development of fast communication interfaces for these platforms. For the GPU and SpiNNaker platforms, we also analysed the power efficiency. A major result was that the host machine that handles data processing for these platforms consumed a significant portion of the total power. This implies that for best performance, the amount of computation and data processing that is carried out on the host machine should be minimised, and as many operations as possible should be offloaded to the devices. Such an approach will also minimise the need for host-device communication and attenuate the impact of bandwidth limitations. These results have been published in Frontiers in Neuroscience (Diamond et al., 2016).
Ongoing work based on these results consists in enabling real-time processing of e-nose signals with the SpiNNaker neuromorphic hardware system. To this end we have designed and built a low-cost gas sensor system that is based on the Arduino microcontroller platform. This system hosts 6 different metal oxide sensors. It can provide readings at intervals as short as tens of milliseconds. Our system enables us to perform our own e-nose experiments, and it serves as a platform to build a real-time interface to the SpiNNaker neuromorphic system and the pattern recognition system that we have implemented thereon.
Publications
Diamond, A, Schmuker, M, Berna, A Z, Trowell, S and Nowotny, Thomas (2016) Classifying continuous, real-time e-nose sensor data using a bio-inspired spiking network modelled on the insect olfactory system. Bioinspiration & Biomimetics, 11(2): 026002.
Schmuker, Michael, Bahr, Viktor and Huerta, Ramón (2016) Exploiting plume structure to decode gas source distance using metal-oxide gas sensors. Sensors and Actuators B: Chemical, 235: 636-646.
Diamond, Alan, Nowotny, Thomas and Schmuker, Michael (2016) Comparing neuromorphic solutions in action: implementing a bio-inspired solution to a benchmark classification task on three parallel-computing platforms. Frontiers in Neuroscience, 9: 491.
Other references
Vergara, A., Fonollosa, J., Mahiques, J., Trincavelli, M., Rulkov, N., and Huerta, R. (2013). On the performance of gas sensor arrays in open sampling systems using Inhibitory Support Vector Machines. Sensors Actuators B Chem. 185, 462–477.
Yavuz, Esin, Turner, James and Nowotny, Thomas (2016) GeNN: a code generation framework for accelerated brain simulations. Scientific Reports, 6: 18854.