Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
© 2012 Boise State University 11© Vishal SaxenaORNL, June 30, 2016
Addressing Challenges in Neuromorphic Computing with Memristive Synapses
Vishal Saxena1, Xinyu Wu1 and Maria Mitkova2
1Analog Mixed-Signal and Photonic IC (AMPIC) Lab 2Nanoionic Materials and Devices Lab
ECE Department, Boise State University, ID
http://lumerink.com
2© Vishal SaxenaORNL, June 30, 2016
Agenda
• Introduction
• Synapse Devices and Challenges
• Mixed-Signal Neural Learning Architecture
• Realizing Multibit Synapses
• Conclusion
3© Vishal SaxenaORNL, June 30, 2016
Reinventing Computing
• Von Neumann computing in the nano-CMOS regime
+ Massive transistor count, but….
– Device variability
– Dark silicon issue
– Not well suited for processing massively parallel data streams and pattern recognition tasks that come easy to a brain.
• US Office of Science and Technology (OSTP) RFI, June 2015:“Create a new type of computer that can proactively interpret and learn from
data, solve unfamiliar problems using what it has learned, and operate with the energy efficiency of the human brain”
DOE, Brain Initiative, AFOSR, DARPA, SRC
4© Vishal SaxenaORNL, June 30, 2016
Brain-Inspired Neuromorphic System
Our Current Research Focus
Source: DARPA SyNAPSE Program
5© Vishal SaxenaORNL, June 30, 2016
Revisiting Biological Inspiration
Dendrites
Soma
Axon
Synapse
Synapse
Synapse
Vspk,in
Vmem
Vspk, out
Vmem
Vspk,in
Vthr
tfire
Vspk, out
t1 t2 t3
Vmem
Neuron
Neuron
Neuron
FromPrevious Neurons
A B C
InhibitiveConnection
Δw
0
PrePost
PrePost
Δt<0 Δt>0
Δt = tpost–tpre
Δt Synaptic placiticity
Vspk,in
Vspk, out
• Emerging understanding of neural computation from Computational Neuroscience
– Hierarchical spiking neural networks with localized learning
• Synapses strengthen or weaken by relative firing times of the pre- and post-neurons
– Plasticity rules: Spike-Timing Dependent Plasticity (STDP)
6© Vishal SaxenaORNL, June 30, 2016
Neuromorphic Computing with RRAMs
Forward spike
Backward spike
Post-synapticneuron
Pre-synapticneuron
ReRAMsynapse
Outp
ut L
aye
r
Input Layer
WTA bus
WTA bus
• To achieve biology-like synapse density, emerging nanoscale devices are under consideration– Resistive RAM (RRAM) or Memristors– Chalcogenide-based Conductive Bridge RAM (CBRAM) devices from Mitkova
Group at Boise State
• Three necessary criterions for the realization of neuromorphic hardware capable of deep learning: (1) non-volatility of the trained weights, (2)massive synaptic density, and (3) ultra-low-power operation.
7© Vishal SaxenaORNL, June 30, 2016
CBRAM I-V Characteristics
I-V Sweep Resistance
8© Vishal SaxenaORNL, June 30, 2016
Stochastic CBRAM BehaviorVTH_Set Distributions RON Distributions
• The CBRAM switching behavior is stochastic• The ‘analog’ synapse value depends upon the electroforming process
• Weak bridge-> more intermediate resistance values• The analog resistance value relaxes in few minutes to hours
9© Vishal SaxenaORNL, June 30, 2016
Spike Timing Dependent Plasticity (STDP)
t
t
t
vp
vn
Vpre
Vpost
tpre tpost
Δt>0
Vnet=Vpost-Vpre
t
t
t
Vpre
Vpost
tpre tpost
Δt<0
Vnet=Vpost-Vpre
Δw or ΔGmr
0
Pre
Post
Pre
Post
Δt<0 Δt>0
Δt = tpost–tpre
Δt
vp
vn
A B
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic
System for Real-World Pattern Recognition”, IEEE Journal on Emerging and
Selected Topics in Circuits and Systems, Accepted.
-4 -2 0 2 4
-0.2
-0.1
0
0.1
0.2
Δt (µs)
ΔG
mr (µ
S)
10© Vishal SaxenaORNL, June 30, 2016
CBRAM Synapse Challenges
• Variability: Device characteristics such as the switching threshold voltages are variable from device to device
– Switching characteristics depend upon the initial ‘forming’ step and compliance current
• Resistive Load: Thousands of devices in parallel present a difficult (low resistance) load to drive by the neuron circuit
• Resolution: Challenging to realize long-term persistence of weights for more than 1-bit resolution in CBRAM devices
11© Vishal SaxenaORNL, June 30, 2016
CMOS Integrate and Fire Neuron
WTA Bus
Cmem
Vthr
Vmem
Vrefr
Mleaky
SW1
SW2
Av
SW3
1
Фd=1
Фf=0
1
SpikeGenerator
Vaxon
Vden
Ф1Ф2
Vleaky Vrest
Фi Фf VtchVmode
Фd
Vwtab
Vrst
Vcpr
Фc
Фd=0
Bus Interface
Phase Controller
• Asynchronous integrated and fire operation
• Accommodates RRAM– Provides current summing node– Sustain a fixed voltage for synapses– Dynamic biasing for large current
drive
• STDP compatible– STDP-compatible spike generator– In-situ STDP learning
• Winner-takes-all bus for effective local decision making
• Single contact point to a synapse
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”, IEEE Journal on Emerging and Selected Topics in Circuits and
Systems, 5(3), 2015.
12© Vishal SaxenaORNL, June 30, 2016
Novel Tri-mode Operation
Cmem
Vthr
Vmem
Vrefr
Mleaky
SW1
SW2
Av
SW3
1
Фd=1
Фf=0
1
SpikeGenerator
Vaxon
Vden
Ф1
Ф2
Vleaky Vrest
Фi Фf VtchVmode
Фd
Vwtab
Vrst
Vcpr
Фc
Фd=0
Bus Interface
Phase Controller
Vrefr
A
Av
VrefrVmem
Vrefr
Vthr
Leaky Integrator
Cmem
Vthr
Vmem
Vrefr
Mleaky
SW1
SW2
Av
SW3
Фd=1
Фd=1
0
Фf=1
SpikeGenerator
Vaxon
Vden
Ф1
Ф2
Vleaky Vrest
Фi Фf VtchVmode
Фd
Vwtab
Vrst
Vcpr
Фc
0
Bus Interface
Phase Controller
Vrefr
BVoltage Buffer
Av
Cmem
Vthr
Vmem
Vrefr
Mleaky
SW1
SW2
Av
SW3
Фd=1
Фd=1
Фf=0
1
Vaxon
Vden
Ф1
Ф2
Vleaky Vrest
Фi Фf VtchVmode
Фd
Vwtab
Vrst
Vcpr
Фc
0
Bus Interface
Phase Controller
Vrefr
CDischarging
Av
Vrefr
Vrefr
SpikeGenerator
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”, IEEE Journal on Emerging and Selected Topics in Circuits and
Systems, 5(3), 2015.
• A first CMOS neuron that is compatible with RRAMs and drives a large load– Several resistive devices in parallel load the neuron
13© Vishal SaxenaORNL, June 30, 2016
Simulation: Output Spikes
Source: X. Wu, V. Saxena and K. Zhu, “A CMOS Spiking Neuron for Dense Memristor-Synapse Connectivity for Brain-Inspired Computing”, International Joint Conference on Neural Networks
(IJCNN), 2015.
Parameter Symbol Unit Min Max Step
Positive amplitude α+ mV 0 360 24
Negative amplitude α- mV 0 360 24
Positive pulse width τ+ ns 48 396 23
Negative pulse width τ- ns 282 1125 56
Negative pulse slope s mV/ns 0.011 0.52 0.034
Output Spikes
α+
α-
τ+
τ-
slope
14© Vishal SaxenaORNL, June 30, 2016
Test Chip interfaced with CBRAM Devices
Source: X. Wu, V. Saxena, K. Zhu and S. Balagopal, “A CMOS Spiking Neuron for Brain-Inspired Neural Networks With Resistive Synapses and In Situ Learning”, IEEE
TCAS-II, 62(11), pp.1088-1092.
• Mixed-signal IC with 180-nm CMOS Neurons connected with CBRAM Synapses• A first RRAM compatible STDP learning Neuron• Power consumed in the neuron is 9.3pJ/spike/synapse (in terms of 10,000
synapses each having around 1MΩ resistance).
Pre-spikeVpre
Post-spikeVpost
Net potential across synapse
Vnet = Vpost - Vpre
+Vt,p
-Vt,m
tail-
tail+
Va-
Va+
A B
Δt
Over threshold potential created by STDP spike-pair
15© Vishal SaxenaORNL, June 30, 2016
Associative Learning Experiment
Source: X. Wu, V. Saxena, K. Zhu and S. Balagopal, “A CMOS Spiking Neuron for Brain-Inspired Neural Networks With Resistive Synapses and In Situ Learning”, IEEE TCAS-II, 62(11),
pp.1088-1092.
Pre-spike
Post-spike
t1
t1+Δt
Post-Neuron
Pre-
Neuron 1
Resistive
Synapses
Outp
ut L
aye
r
Input Layer
Neurons
IFN
#1
IFN
#3
Pre-
Neuron 2
IFN
#2
t2
Pre-Neuron 1(IFN #1)
"Sight of Food"
Pre-Neuron 2 (IFN #2)
"Sound of Bell"
Post-Neuron (IFN #3)
"Salivation"
Time (ms)
16© Vishal SaxenaORNL, June 30, 2016
Dynamic Resistive Load Driving
0 2 4 6 8 10
0
400
800
1200
1600
Time (us)
Tota
l C
urr
ent (u
A)
10
100
1,000
10,000 1,000 100 10 1
Cu
rre
nt
(µ
A)
Number of Synapses
SynapsesDrivingBaseline
13µA56µA
1,000 Synapses100 Synapses
10,000 Synapses
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”, IEEE Journal on Emerging and Selected Topics in Circuits and
Systems, 5(3), 2015.
97% Efficiency@ driving 10,000 x
1MΩ synapses
17© Vishal SaxenaORNL, June 30, 2016
Spike-based Pattern Recognition Circuit
...
Image i×j Input Network
i×j Input Neuron Matrix
i×j×n Synaptic Network
n Output Neurons
WTA
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”, IEEE Journal on Emerging and Selected Topics in Circuits and
Systems, 5(3), 2015, in press.
18© Vishal SaxenaORNL, June 30, 2016
Asynchronous WTA Bus Interface
Inhibitory connectionExcitatory connection
Explicit one-one inhibitory connections Implicit inhibition via shared WTA bus
ΦfD
S
Vwtab
Vcpr
Vtch
VmodeΦd
Vrst
Φc
Q
Vdd
DelayTri-state Buffer
Inhibitory connectionExcitatory connection
Explicit one-one inhibitory connections Implicit inhibition via shared WTA bus
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”, IEEE Journal on Emerging and Selected Topics in Circuits and
Systems, Accepted.
WTA Bus
Cmem
Vthr
Vmem
Vrefr
Mleaky
SW1
SW2
Av
SW3
1
Фd=1
Фf=0
1
SpikeGenerator
Vaxon
Vden
Ф1Ф2
Vleaky Vrest
Фi Фf VtchVmode
Фd
Vwtab
Vrst
Vcpr
Фc
Фd=0
Bus Interface
Phase Controller
19© Vishal SaxenaORNL, June 30, 2016
Demo: Handwrite Digits Classification
0 100 200 300 400
10
20
30
40
50
60
Input Spike Trains
Neu
ron
#
Time (µs)
0 100 200 300 400
0
1
2
7
Output Spike Trains without Learning
Neu
ron
#
Time (µs)
0 100 200 300 400
0
1
2
7
Output Spikes Train with Learning
Neu
ron
#
Time (µs)
0 1 2 7
Source: X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”, IEEE Journal on Emerging and Selected Topics
in Circuits and Systems, 5(3), 2015.
Conductance Evolution in Synapses (µS)
10
20
30
40
50
0 1094 2189 3283 4377 5471 6566 7660Time (µs)
0
First 20
Train
(original sequence)
Test
(class-by-class)
20© Vishal SaxenaORNL, June 30, 2016
Parallel Multibit Synapses
• Emerging spike-based deep learning architectures require at least 4-bits of synaptic resolution.
• The devices fundamentally have bistable nonvolatile states with stochastic switching behavior
• Recent approach: Use several of these device in parallel – Potentially multiple resistance
states due to variation in individual threshold voltages
Output Neuron
Input Layer
Current integration
Compound Multibit
Synapse
Forward
spike
Backward
spikes
21© Vishal SaxenaORNL, June 30, 2016
Multibit Synapses: Dendritic Processing
• Dendritic processing to include sensitivity to time-overlap between the pre- and post-synaptic spikes
• Shift in time or pulse voltage in the back-propagating spikes
• Simulations exhibit remarkable STDP learning behavior; experiments underway with physical devices
X. Wu and V. Saxena, “Realizing Multibit Synapses with Bistable RRAM Devices”, Unpublished result.
22© Vishal SaxenaORNL, June 30, 2016
Conclusion
• Nanoscale emerging memory devices present opportunity for realizing non von Neumann computing
• Mixed-Signal architectures for spiking neural networks can realize a versatile neuromorphic computing motif
• Further need to characterize device capabilities and limitations for large scale integration
• Algorithm/architecture development to be driven ground up by emerging device capabilities
23© Vishal SaxenaORNL, June 30, 2016
References1. X. Wu, V. Saxena and K. Zhu, “Homogeneous Spiking Neuromorphic System for Real-World Pattern Recognition”,
IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 5(3), 2015, in press.
2. X. Wu, V. Saxena and K. Zhu, “A CMOS Spiking Neuron for Dense Memristor-Synapse Connectivity for Brain-Inspired Computing”, International Joint Conference on Neural Networks (IJCNN), 2015, Accepted.
3. X. Wu, V. Saxena and K. A. Campbell, “Energy-efficient STDP-based learning circuits with memristor synapses”, SPIE Machine Intelligence and Bio-inspired Computation: Theory and Applications VIII, 2014.
4. X. Wu, V. Saxena and K. Zhu, “A CMOS Spiking Neuron for Brain-Inspired Neural Networks with Resistive Synapses and In-Situ Learning”, IEEE Transactions on Circuits and Systems II: Express Briefs, In final review.
5. X. Wu and V. Saxena, “Associative Learning on CMOS-Memristor Spiking Neuromorphic Chip”, IEEE Transactions on Nanotechnology, In 2nd round review.
24© Vishal SaxenaORNL, June 30, 2016
Questions?