Limits...
Implementation of olfactory bulb glomerular-layer computations in a digital neurosynaptic core.

Imam N, Cleland TA, Manohar R, Merolla PA, Arthur JV, Akopyan F, Modha DS - Front Neurosci (2012)

Bottom Line: Our system is based on a digital neuromorphic chip consisting of 256 leaky-integrate-and-fire neurons, 1024 × 256 crossbar synapses, and address-event representation communication circuits.The neural circuits configured in the chip reflect established connections among mitral cells, periglomerular cells, external tufted cells, and superficial short-axon cells within the olfactory bulb, and accept input from convergent sets of sensors configured as olfactory sensory neurons.Our circuits, consuming only 45 pJ of active power per spike with a power supply of 0.85 V, can be used as the first stage of processing in low-power artificial chemical sensing devices inspired by natural olfactory systems.

View Article: PubMed Central - PubMed

Affiliation: Computer Systems Lab, Department of Electrical and Computer Engineering, Cornell University Ithaca, NY, USA.

ABSTRACT
We present a biomimetic system that captures essential functional properties of the glomerular layer of the mammalian olfactory bulb, specifically including its capacity to decorrelate similar odor representations without foreknowledge of the statistical distributions of analyte features. Our system is based on a digital neuromorphic chip consisting of 256 leaky-integrate-and-fire neurons, 1024 × 256 crossbar synapses, and address-event representation communication circuits. The neural circuits configured in the chip reflect established connections among mitral cells, periglomerular cells, external tufted cells, and superficial short-axon cells within the olfactory bulb, and accept input from convergent sets of sensors configured as olfactory sensory neurons. This configuration generates functional transformations comparable to those observed in the glomerular layer of the mammalian olfactory bulb. Our circuits, consuming only 45 pJ of active power per spike with a power supply of 0.85 V, can be used as the first stage of processing in low-power artificial chemical sensing devices inspired by natural olfactory systems.

No MeSH data available.


Related in: MedlinePlus

The sparse connectivity of the ET/sSA network on the chip is functionally equivalent to a fully connected network, but carries out global normalization at a reduced energy cost. The abscissa is a measure of the density of sSA connections, bounded by the extremes of a fully isolated network at x = 0 and an all-to-all interconnected network at x = 48. The primary ordinate (left) denotes the coefficient of variation (CV% – the standard deviation as a percentage of the mean) of sSA activity across the sSA population. Low CVs indicate high uniformity in sSA activity across the network. The secondary ordinate (right) denotes the energy consumed in the chip (at its present scale) by updating all synaptic inputs when sSA neurons spike. Denser interconnectivity requires more energy. The dotted vertical line denotes the density of sSA innervation presently implemented in the chip. This reasonably optimized solution corresponds directly to that which appears to be implemented in mouse and rat olfactory bulbs (cf. Figure 5B in Cleland et al., 2007).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC3368244&req=5

Figure 11: The sparse connectivity of the ET/sSA network on the chip is functionally equivalent to a fully connected network, but carries out global normalization at a reduced energy cost. The abscissa is a measure of the density of sSA connections, bounded by the extremes of a fully isolated network at x = 0 and an all-to-all interconnected network at x = 48. The primary ordinate (left) denotes the coefficient of variation (CV% – the standard deviation as a percentage of the mean) of sSA activity across the sSA population. Low CVs indicate high uniformity in sSA activity across the network. The secondary ordinate (right) denotes the energy consumed in the chip (at its present scale) by updating all synaptic inputs when sSA neurons spike. Denser interconnectivity requires more energy. The dotted vertical line denotes the density of sSA innervation presently implemented in the chip. This reasonably optimized solution corresponds directly to that which appears to be implemented in mouse and rat olfactory bulbs (cf. Figure 5B in Cleland et al., 2007).

Mentions: where Pij is the probability of a synaptic connection between glomerular column i and sSA j, Po is a baseline probability factor, fgauss represents a gaussian distribution (parameterized by variance R), and dij is the distance between the glomerular column and the sSA cell. By varying Po and R, we changed the synaptic density and the projection distance of each sSA cell. An increase in either of these parameters increased the amount of excitation that each glomerular column receives from sSA cells. In order to observe only the effects of changing the connection distribution of the network and not changes in its overall strength, we reduced individual synaptic weights in proportion to the increase in connectivity. The coefficient of variation (CV) of sSA activity across the chip provides a measure of how uniform the network activity is, and consequently, the efficacy of the normalization process. We performed this analysis for several different odor-activated sensor patterns and averaged the results. When the average number of sSA inputs per glomerulus is low (0 in the extreme case of a completely isolated glomerulus), the CV is large, reflecting the heterogenous activation levels of different glomeruli. The CV converges rapidly to an asymptotic minimum as the density of sSA connections is increased (Figure 11; also see Cleland et al., 2007) The connectivity profile depicted by the vertical dotted line in Figure 11 (representing an average of 10 sSA inputs per glomerulus) substantially reduced energy consumption while generating approximately the same CV for sSA activity as that achieved by a fully connected ET/sSA network. Good normalization results were obtained for sSA densities as low as four inputs per glomerulus. These results demonstrate that a small-world network on the chip can achieve close to the maximum normalization quality using up to 10 times less energy compared to an all-to-all network. As we scale the system (through multiple cores), this effect will become increasingly significant, since all-to-all ET/sSA connectivity at larger scales would consume an increasingly disproportionate share of chip resources.


Implementation of olfactory bulb glomerular-layer computations in a digital neurosynaptic core.

Imam N, Cleland TA, Manohar R, Merolla PA, Arthur JV, Akopyan F, Modha DS - Front Neurosci (2012)

The sparse connectivity of the ET/sSA network on the chip is functionally equivalent to a fully connected network, but carries out global normalization at a reduced energy cost. The abscissa is a measure of the density of sSA connections, bounded by the extremes of a fully isolated network at x = 0 and an all-to-all interconnected network at x = 48. The primary ordinate (left) denotes the coefficient of variation (CV% – the standard deviation as a percentage of the mean) of sSA activity across the sSA population. Low CVs indicate high uniformity in sSA activity across the network. The secondary ordinate (right) denotes the energy consumed in the chip (at its present scale) by updating all synaptic inputs when sSA neurons spike. Denser interconnectivity requires more energy. The dotted vertical line denotes the density of sSA innervation presently implemented in the chip. This reasonably optimized solution corresponds directly to that which appears to be implemented in mouse and rat olfactory bulbs (cf. Figure 5B in Cleland et al., 2007).
© Copyright Policy - open-access
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC3368244&req=5

Figure 11: The sparse connectivity of the ET/sSA network on the chip is functionally equivalent to a fully connected network, but carries out global normalization at a reduced energy cost. The abscissa is a measure of the density of sSA connections, bounded by the extremes of a fully isolated network at x = 0 and an all-to-all interconnected network at x = 48. The primary ordinate (left) denotes the coefficient of variation (CV% – the standard deviation as a percentage of the mean) of sSA activity across the sSA population. Low CVs indicate high uniformity in sSA activity across the network. The secondary ordinate (right) denotes the energy consumed in the chip (at its present scale) by updating all synaptic inputs when sSA neurons spike. Denser interconnectivity requires more energy. The dotted vertical line denotes the density of sSA innervation presently implemented in the chip. This reasonably optimized solution corresponds directly to that which appears to be implemented in mouse and rat olfactory bulbs (cf. Figure 5B in Cleland et al., 2007).
Mentions: where Pij is the probability of a synaptic connection between glomerular column i and sSA j, Po is a baseline probability factor, fgauss represents a gaussian distribution (parameterized by variance R), and dij is the distance between the glomerular column and the sSA cell. By varying Po and R, we changed the synaptic density and the projection distance of each sSA cell. An increase in either of these parameters increased the amount of excitation that each glomerular column receives from sSA cells. In order to observe only the effects of changing the connection distribution of the network and not changes in its overall strength, we reduced individual synaptic weights in proportion to the increase in connectivity. The coefficient of variation (CV) of sSA activity across the chip provides a measure of how uniform the network activity is, and consequently, the efficacy of the normalization process. We performed this analysis for several different odor-activated sensor patterns and averaged the results. When the average number of sSA inputs per glomerulus is low (0 in the extreme case of a completely isolated glomerulus), the CV is large, reflecting the heterogenous activation levels of different glomeruli. The CV converges rapidly to an asymptotic minimum as the density of sSA connections is increased (Figure 11; also see Cleland et al., 2007) The connectivity profile depicted by the vertical dotted line in Figure 11 (representing an average of 10 sSA inputs per glomerulus) substantially reduced energy consumption while generating approximately the same CV for sSA activity as that achieved by a fully connected ET/sSA network. Good normalization results were obtained for sSA densities as low as four inputs per glomerulus. These results demonstrate that a small-world network on the chip can achieve close to the maximum normalization quality using up to 10 times less energy compared to an all-to-all network. As we scale the system (through multiple cores), this effect will become increasingly significant, since all-to-all ET/sSA connectivity at larger scales would consume an increasingly disproportionate share of chip resources.

Bottom Line: Our system is based on a digital neuromorphic chip consisting of 256 leaky-integrate-and-fire neurons, 1024 × 256 crossbar synapses, and address-event representation communication circuits.The neural circuits configured in the chip reflect established connections among mitral cells, periglomerular cells, external tufted cells, and superficial short-axon cells within the olfactory bulb, and accept input from convergent sets of sensors configured as olfactory sensory neurons.Our circuits, consuming only 45 pJ of active power per spike with a power supply of 0.85 V, can be used as the first stage of processing in low-power artificial chemical sensing devices inspired by natural olfactory systems.

View Article: PubMed Central - PubMed

Affiliation: Computer Systems Lab, Department of Electrical and Computer Engineering, Cornell University Ithaca, NY, USA.

ABSTRACT
We present a biomimetic system that captures essential functional properties of the glomerular layer of the mammalian olfactory bulb, specifically including its capacity to decorrelate similar odor representations without foreknowledge of the statistical distributions of analyte features. Our system is based on a digital neuromorphic chip consisting of 256 leaky-integrate-and-fire neurons, 1024 × 256 crossbar synapses, and address-event representation communication circuits. The neural circuits configured in the chip reflect established connections among mitral cells, periglomerular cells, external tufted cells, and superficial short-axon cells within the olfactory bulb, and accept input from convergent sets of sensors configured as olfactory sensory neurons. This configuration generates functional transformations comparable to those observed in the glomerular layer of the mammalian olfactory bulb. Our circuits, consuming only 45 pJ of active power per spike with a power supply of 0.85 V, can be used as the first stage of processing in low-power artificial chemical sensing devices inspired by natural olfactory systems.

No MeSH data available.


Related in: MedlinePlus