2cm Emergent auditory feature tuning in a real-time neuromorphic VLSI …/file/Elisabetta... ·...
Transcript of 2cm Emergent auditory feature tuning in a real-time neuromorphic VLSI …/file/Elisabetta... ·...
Emergent auditory feature tuning in a real-timeneuromorphic VLSI system
Elisabetta Chicca
Cognitive Interaction Technology Center of ExcellenceUniversity of Bielefeld
STINT Workshop 2012
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 1
Neuromorphic Engineering
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 2
Neural computation→ neurotechnologies
The term neuromorphic was coined by Carver Mead, in the late 1980s to describeVery-Large-Scale Integration (VLSI) systems containing sub-threshold analog circuitsthat mimic neuro-biological architectures present in the nervous system.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 3
Neuromorphic circuits
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 4
Building blocksThe biological neuron
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 5
Integrate-and-Fire neuron
0
10
20
30
Neuro
n
0 1 2Time (s)
0 1 20
10
20
30
Neuro
n
Time (s)
G. Indiveri, et al. “A VLSI array of low-power spiking neurons and bistable synapses with spike-timing dependent plasticity”, IEEE Transactionson Neural Networks, 2006
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 6
Integrate-and-Fire neuron
0
10
20
30
Ne
uro
n
0 1 2Time (s)
0 1 20
10
20
30
Ne
uro
n
Time (s)0.5 0.6 0.7 0.8
100
101
102
−Vgs
(V)
<f>
(H
z)
Vrfr
G. Indiveri, et al. “A VLSI array of low-power spiking neurons and bistable synapses with spike-timing dependent plasticity”, IEEE Transactionson Neural Networks, 2006
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 6
Synaptic dynamicsImpulse response and short-term depression
C. Bartolozzi and G. Indiveri, “Synaptic dynamics in analog VLSI”, Neural Computation, 2007.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 7
Synaptic dynamicsLearning
VL
θ
VH
wV
mem
0 0.1 0.2 0.3 0.4
pre
Time(s)
VL
θ
VH
wV
mem
0 0.1 0.2 0.3 0.4
pre
Time(s)
Mitra et al., “Real-time classification of complex patterns using spike-based learning in neuromorphic VLSI”, IEEE Transaction on BiomedicalCircuits and Systems, 2009.Giulioni et al.,“Classification of correlated patterns with a configurable analog VLSI neural network of spiking neurons and self-regulating plasticsynapses.”, Neural Computation, 2009.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 8
Neuromorphic systems
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 9
Address Event Representation (AER)
3
2
1
Inputs
Encode Decode
Address Event Bus
Source
Chip
3
2
1
Outputs
3 2 1 2 1 32
Destination
Chip
Action Potential
Address-Event
representation of
action potential
R. S. Deiss, R. J. Douglas, and A. M. Whatley “A Pulse–Coded Communications Infrastructure for NeuromorphicSystems” in Pulsed Neural Networs, 1998.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 10
Multi-chip setup
PC workstation(A)
AER mapper
AER eventsare transmittedon a serial bus
CHIP-1 (B1)
SATA USB
CHIP-3 (C)
CHIP-2 (B2)
Sheik et al. “Emergent auditory feature tuning in a real-time neuromorphic VLSI system” Frontiers in Neuromorphic Engineering, 2012.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 11
Emergent auditory feature tuning
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 12
Emergent auditory feature tuningFull neural network diagram
Martin Coath et al. “Emergent Feature Sensitivity in a Model of the Auditory Thalamocortical System” Advances in ExperimentalMedicine and Biology, 2011.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 13
Emergent auditory feature tuningHardware neural network diagram
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
Sheik et al. “Emergent auditory feature tuning in a real-time neuromorphic VLSI system” Frontiers in Neuromorphic Engineering, 2012.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 14
Implementing delays exploiting device mismatch
Sheik et al. “Exploiting Device Mismatch in Neuromorphic VLSI Systems to Implement Axonal Delays” International Joint Conference on NeuralNetworks, 2012.
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 15
Emergent auditory feature tuningInput stimuli
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 16
Emergent auditory feature tuningSynaptic Matrix after learning
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 17
Emergent auditory feature tuningSynaptic Matrix after learning
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 18
Emergent auditory feature tuningTuning curve
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 19
Emergent auditory feature tuningTuning curve
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 20
Conclusions
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 21
Animal noises and argo floats
Lars Kindermann, “Strange Sounds of theSouthern Ocean”
Ocean Acoustics Lab, Alfred Wegener Institute,
DE
http://www.whoi.edu/
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 22
Acknowledgments
Sadique Sheik§, Martin Coath�,Giacomo Indiveri§, Susan Denham�, Thomas Wennekers◦
Chiara Bartolozzi\, Emre Neftci§, Daniel Fasnacht§, Fabio Stefanini§, andRodney Douglas§
Funding: ICT-231168-SCANDLE “acoustic SCene ANalysis for DetectingLiving Entities” EU grant, DAISY (FP6-2005-015803) EU grant, and Cluster ofExcellence 277 (CITEC, Bielefeld University).
§ Institute of Neuroinformatics, University and ETH Zurich, CH� School of Psychology, University of Plymouth, UK◦ School of Computing and Mathematics, University of Plymouth, UK\ Italian Institute of Technology (IIT), IT
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 23
Van Rossum distanceTuning curve
A
B2B1
= Excitatory = Excitatory STDP = Inhibitory
C
A
B2B1
C
A
B2B1
C
Given two spike trains s1 and s2, the Van Rossum distance is defined as:
D(s1,s2) =
√∫∞
−∞
[g ∗ s1−g ∗ s2]2dt
where g = g(t;τc) is a smoothing function (e.g., a decaying exponential) withtime constant τc .
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 24
The VLSI AER transceiver chip (IFSLWTA)
Technology: AMS 0.35µmTotal area: 3.94mm×2.54mmCore area: 2.6mm×1.9mmNeurons: 128 (124 exc. + 4 inh.)Synapses: 32×128Dendritic tree multiplexer: 32x128 | 64x64 | ... | 4096x1|
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 25
The 2D VLSI AER transceiver chip (2DIFWTA)
Technology: AMS 0.35µmArea: 5.14mm×2.94mmNeurons: 2048 (32×64)AER Synapses: 2048×3Local Synapses: 2048×11
E. Chicca(CITEC) Neuromorphic auditory features STINT Workshop 26