Fast electronic noses through spiking neuromorphic networks Prof. Thomas Nowotny CCNR, Informatics, Sussex Neuroscience, University of Sussex Efutures.

Fast electronic noses through spiking neuromorphic networks

Prof. Thomas NowotnyCCNR, Informatics, Sussex Neuroscience,

University of Sussex

Efutures Community EventBritish Museum London, 04-12-2013

The problemEnoses are slow

Chemical sensors are much slower than animals’ sensors

The analysis of sensor data is “slow”:• Based on entire measurement• Done “offline”

• Animals make decisions long before their receptors reach equilibrium

• Decisions are made “online” with a continuous input stream

• Use biomimetic spiking neural networks

• Simulating SNN is slow(ish) • Use neuromorphic hardware to

accelerate to hyper-realtime

FOX enose

MOx sensors(Figaro)

FOX enose system

O2+Analyte

I

Heater

Substrate

Metal Oxide

The data is theResistance Change

Two different sensor technologies

I

Heater

SNO2

Classical SNO2 Sensors

I

Heater

CTO

The data is theResistance Change

Zeolite-coated CTO Sensors

O2+AnalyteO2+Analyte Zeolite

coating

SubstrateSubstrate

Example data

Hexanol Octenol

Zeolite CTO sensors

SNO2 sensors SNO2 sensors

Zeolite CTO sensors

(here R0 was subtracted)

Time (s)Time (s)

Rela

tive

resp

onse

(au)

Rela

tive

resp

onse

(au)

Faster Features

e.g. A. Z. Berna et al. 2011 ISOEN Conference, New York

R/R0

EMAmax for 3 timescales

Traditional: Steady State

Faster: Transients

Traditional approach: Measure steady states

activation Use discriminant analysis

and/or machine learning methods

Bio-mimetic online approach: Use spiking neural network Make “guesses” continuously

in real time Use neuromorphic systems to

make this viable

Models:

Pfeil et al., Frontiers in Neuroscience 2013Huerta et al., Neural Computation 2009

Implementation: GeNN GPU Kit and Leicester FPGA Kit

GPU

NVIDIA Tesla

FPGA

Xilinx Virtex

Guerrero-Rivera et al., Programmable Logic Construction Kits for Hyper-Real-TimeNeuronal Modeling. Neural Computation 18, 2651–2679 (2006)

Nowotny et al., GPU enhanced Neuronal Networks (GeNN), BMC Neuroscience 2011, 12(Suppl 1):P239.http://genn.sourceforge.net

Project plan

WP1: Objective: Develop and verify a spiking network prototype for rapid analysis of chemosensor signals. (month 1-5)

Tasks: Implement a GPU‐accelerated spiking network. Tune it for performance on the basis of e‐nose data sets. Benchmark the performance against conventional state‐of‐the‐art approaches

Outcome: A GPU‐accelerated spiking network for e‐nose signal analysis.

WP2: Objective: Port the network to neuromorphic hardware. (month 6-7)

Tasks: Implement the network using the neuromorphic kit from Leicester (Tim C. Pearce) Verify that the network’s performance on hardware is at level with the software

implementation.

Outcome: Hardware implementation of the spiking e‐nose network.

Future Perspectives

Porting to SPIKEY (Karlheinz Meier, Heidelberg) Scaling the classifiers to HiCANN and wafer-size system Exploring implementations for SpiNNaker (Steve Furber,

Manchester)

M Schmuker has 2 year Marie Curie Fellowship from September 2014.

We (M Schmuker & T Nowotny) have applied for HBP funding to further pursue this.

The Team

PIs ConsultantTim Pearce:

Spiking NN on FPGA

Bio-mimetic classification model 3

Russell Binions/ Amalia Berna:

Sensor technology and Enose data

Thomas Nowotny:(overall lead)

Bio-mimetic classification model 2

Spiking NN on GPU

Michael Schmuker:

Biomimetic classification model 1

Spiking NN on neuromorphic hardware (SPIKEY)

ResearcherInterviews: next week

Acknowledgments

More info: email [email protected]