neuromorphic computing with nanocrossbar circuits · drumpf twitterbot learns to imitate trump via...
TRANSCRIPT
Neuromorphic Computing with NanoCrossbar Circuits
Dmitri StrukovUniversity of California at Santa Barbara
Acknowledgments:
Research group: G. Adam, B. Chakrabarti, X. Guo, B. Hoskins, F. Merrikh Bayat, M. Prezioso
Collaborators: P. Auroux, J. Edwards, M. Graziano (BAE), I. Kataeva (DENSO), K. K. Likharev (SBU), N. Do (SST), L. Sengupta (NGC)
Funding support: AFOSR MURI, ARO, DARPA UPSIDE, DENSO CORP., NSF
1
NanoXbar Workshop, July 2016
2
RECENT SURGE OF A.I.(you could not avoid the buzz…)
Drumpf Twitterbot learns to imitate Trump via deep-learning algorithm
Donald TrumpBot@thetrumpbot
“OK, it's amazing right now with ISIS, I tell you what? I don't want them to vote, the worst very social people. I love me”
NanoXbar Workshop, July 2016
3
PATTERN CLASSIFICATION
IN CONVOLUTIONAL (A.K.A. “DEEP”) NETWORKS
Input
1@32x32
Layer 1
6@28x28
Layer 2
6@14x14
Layer 3
12@10x10Layer 4
12@5x5
Layer 5
100@1x1
10
5x5convolution
5x5convolution
5x5convolution2x2
pooling/sub-sampling
2x2pooling/sub-sampling
Layer 6: 10
- MLP with limited bio-inspired connectivity
- the best method for hand-writing recognition
- used by NCR for check reading machines
- used by Microsoft for OCR
- 0.62% error on the MNIST set
Y. LeCun (2007)
Example:
MNIST set (60,000-image database)
NanoXbar Workshop, July 2016
4
PATTERN CLASSIFICATION
IN CONVOLUTIONAL (A.K.A. “DEEP”) NETWORKS
Input
1@32x32
Layer 1
6@28x28
Layer 2
6@14x14
Layer 3
12@10x10Layer 4
12@5x5
Layer 5
100@1x1
10
5x5convolution
5x5convolution
5x5convolution2x2
pooling/sub-sampling
2x2pooling/sub-sampling
Layer 6: 10
- supervised error
backpropagation training
- weights determine functionality
- neurons learn “features”
- most computationally intensive
operation: dot product
- 4 to 5-bit synapses (weights)
accuracy sufficient
Y. LeCun (2007)
Example:
NanoXbar Workshop, July 2016
j
n
j
j xwo
1
]tanh[oy
inputs
weights
activation function
weighted sum
Single layer perceptron:
5
U. TORONTO’S NETWORK
- 650,000 neurons, 0.63109 synapses
- Image Net LSVRS-2010 benchmark set
- 1.2 M images; 1,000 classes
- error rates: top-1 37.5%, top-5 17%
NanoXbar Workshop, July 2016
Bottleneck: massive number of dot
product (vector-by-matrix) computations
between analog inputs and analog
(fixed) weights
Google’s Tensor Processing Unit
DIGITAL CIRCUITS FOR DEEP LEARNING
Movidius’s fanthom
15 inferences /sec @ 16-bit FP precision for ImageNet@ <2W
Nvidia’s Pascal
21 TFLOPS for deep learning performance
- Proposed by Carver Mead and his
students 25+ years ago
- Exact analog-domain dot-product due to
Ohm’s and Kirchhoff's law
- No need to waste energy on memory
bits movement (in-memory computing)
- Major challenge: adjustable cross-point
devices
- Two very promising recent options:
- Custom-built metal-oxide memristors
- Redesigned commercial NOR flash
- Other (not discussed) options: phase
change, ferroelectric, and magnetic
devices
7
NanoXbar Workshop, July 2016
ANALOG VECTOR-BY-MATRIX COMPUTATION
Strukov et al., DRC’16
MEMRISTORS
8
modulation of defect profile
electrolyte (ECM) metal oxide (VCM)
ON state
OFF state
redoxreaction
primary mechanism:
Pt
TiO2-x
Pt
Pt
Ag2S
Ag
-1.0 0 1.0
-1.0
0
1.0
Voltage (V)
Cu
rren
t (m
A) =
+Vswitch-Vswitch
Alibart et al., Nature Comm, 2013
S
S
A
V
Analog switching: Any state between ON and OFF Strongly (superexp) nonlinear switching dynamics Gray area = no change Memory state defined as current measured within gray area
25 nm Au / 15 nm Pt top electrode
5 nm Ti / 25 nm Pt bottom electrode
e-beam patterned Pt protrusion
30 nm TiO2-x
20 nm
Typical I-V for Pt/TiO2-x/Pt devices Two major types of memristors
NanoXbar Workshop, July 2016
0 1000 2000 3000
1E-5
1E-4
7A
120A
60A
30A
Curr
ent @
-200m
V (
A)
Pulse Number
15A
Increase WeightDecrease Weight
Stand-by (Read only)
2006 2007 2008 2009 2010 20111,000
100,000
1E7
1E9
1E11
1E13
Fujitsu Labs
Panasonic Corp.
SAIT
En
dru
an
ce
(cycle
s)
Year
HP Labs
several groups
Kawahara et al. Panasonic, 2012
Alibart et al, Nanotechnology 23 074508, 2012
Schindler, PhD Thesis, 2009
Govoreanu,et all IEDM, 2012
Strachan et al, Nanotechnology 22505402 2011
ON
OFF
Torrezan et al, Nanotechnology 22 485203 2011
J. Yang, DBS, and D. Stewart Nature Nano 8 13-24 (2013)
4) switching speed
2) endurance
5) retention 6) ON/OFFcurrent ratio
7) OFF stateresistance
8) I-V nonlinearity
3) reciprocal switching energy
10) density
1) reproducibility
9) number of states
memorylogicneuro
storage
demonstrated
9
STATE-OF-THE-ART PERFORMANCE
NanoXbar Workshop, July 2016
10
10 100
10
100
Aver
age
R (k
)
Setpoint R (k)
10 12 14 16 18 20 22
10
12
14
16
18
20
22
Avera
ge R
(k
)
Setpoint R (k)
Target resistance @ 0.1V (kΩ)10 100
100
10
222018161412
10
10 12 14 16 18 20 22
Ave
rage
res
ista
nce
@ 0
.1V
(kΩ
)
-2.0-1.5-1.0-0.5 0.0 0.5 1.0 1.5 2.0
-600
-500
-400
-300
-200
-100
0
100
200
300
Cu
rre
nt (
A)
Voltage (V)
Reset
SetForm
SiO2/Si
Pt (60 nm)
TiO2-x (30 nm)
Ti (15 nm)
Al2O3 (4 nm)
Ta (5 nm)
Pt (60 nm)
VW-
VRVW
+
Voltage (V)-2.0 -1.5 -1.0 -0.5 0.0 0.5 1.0 1.5 2.0
300
200
100
0
-100
-200
-300
-400
-500
-600C
urr
ent
(μA
)
-1.5 -1.0 -0.5 0.0 0.5 1.00
5
10
15
Cou
nts
Threshold Voltage (V)
1.7 1.8 1.9 2.0 2.10
10
20
30
40
Cou
nts
Forming voltage (V)
0.3 0.4 0.5 0.60
5
10
15
Cou
nts
Conductance @0.1V (S)
Reset Set
Switching threshold voltage (V)
Forming voltage (V)
Conductance @ 0.1V (μS)
Co
un
tC
oun
tC
oun
t
15
10
5
0
40302010
0
15
10
5
0-1.5 -1.0 -0.5 0.0 0.5 1.0 1.5
1.7 1.8 1.9 2.0 2.1
0.3 0.4 0.5 0.6
Re
sis
tan
ce
@0
.1V
(Ω
)
105
104
Time (s)0 20 40 60 80 100 120 140 160 180
Major features
- 0T1R- 200 nm wide lines - Al2O3 and TiO2-x by
sputtering - Very uniform (~17%)
norm. RMS of [email protected] for 8x10 virgin array
- >500K stress pulses without much degradation
PASSIVE MEMRISTIVE CROSSBAR CIRCUIT Crossbar circuit
2 μm
Typical I-Vs Statisticsas
fabricated
forming
switching threshold
Analog properties and state tuning
M. Prezioso et al., Nature May 2015 M. Prezioso et al., IEDM’15NanoXbar Workshop, July 2016
CLASSIFIER OPERATION (INFERENCE)
- Neurons functionality (opamp) is emulated in software- Differential pair of memristors per weight
differential pair
A A A A
V1
V2
V10
G1,1+ G1,1
-
G1,2+ G1,2
-
G1,10+ G1,10
-
G3,1+ G3,1
-
G3,2+ G3,2
-
G3,10+ G3,10
-
bias
pattern (“z”)
V=
I1+ I1
- I2+ I2
- I3+ I3
-
+VR
-VR
-VR
+VR
+VR
+VR
-VR
-VR
+VR
-VR
11
NanoXbar Workshop, July 2016
M. Prezioso et al., Nature May 2015M. Prezioso et al., IEDM’15
Σ
V10
V1
V2
bias
input
neurons
output neurons
V1 V4 V7
V2 V5 V8
V3 V6 V9
V10
3×3 binary
image
W1,1
W3,10
𝑓1Σ
weights
Σ 𝑓3
𝐼1
𝐼3𝑓2
𝐼2
CLASSIFIER IN-SITU TRAINING (WEIGHT UPDATE)
differential pair
A A A A
V1
V2
V10
G1,1+ G1,1
-
G1,2+ G1,2
-
G1,10+ G1,10
-
G3,1+ G3,1
-
G3,2+ G3,2
-
G3,10+ G3,10
-
Analog vector-by-matrix multiplier
- Half-biasing technique- One column at a time (fully parallel
possible with stochastic training mode)
- Neurons functionality (opamp) is emulated in software- Differential pair of memristors per weight
bias
pattern (“z”)
V=
I1+ I1
- I2+ I2
- I3+ I3
-
+VR
-VR
-VR
+VR
+VR
+VR
-VR
-VR
+VR
-VR
step 1: set G+
step 2: reset G+
step 3: reset G-
step 4: set G-
0 0 0 00
0
0
0
0
0
0 0 0 00
-VW+/2
00
0
0
0
-VW+/2
-VW+/2
-VW+/2
-VW+/2
+VW+/2
+VW+/2
+VW+/2
+VW+/2
+VW+/2
-VW+/2
+VW+/2
sgn[ΔW]=
calculate
fi (n) =
calculate
𝐼𝑖(n) =
calculateset
n = 1
Δij = 0
desired class fi(g)(n)
last patter
n?
no
yes
update
weights
Vj(n)
n = n +1
next epoch
end of
epoch
no
last patter
n?
yesupdate
weights
n = n +1
stochastic
batch
J
j
jijVW1
iItanh
nVnn jiij
)(
)(
nIIi
i
g
iidI
dfnfnfn
training set:
𝑉𝑗 𝑛 , 𝑓𝑖(g)(𝑛)𝑛=1
𝑁
initializeWij
sgn[ΔW]=
sgn[δi(n)Vj (n)]0
0
00
0
0 00
0
00
0
0
0 00
-VW+/2
-VW+/2
-VW+/2
-VW+/2
-VW+/2
+VW+/2+VW
+/2+VW+/2
+VW+/2
+VW+/2
+VW+/2
+VW+/2
+VW+/2
-VW+/2 -VW
+/2-VW+/2
Weight updatestochastic
batch
12
NanoXbar Workshop, July 2016
M. Prezioso et al., Nature May 2015M. Prezioso et al., IEDM’15
EXPERIMENTAL RESULTS
0
20
40
60
80
100
0 10 20 30 40 5005
1015202530
Test set
26% misclassification
# m
iscla
ssifie
d p
attern
s
Training set
Epoch
Batch Manhattan rule in-situ training:
Trained on the original training set
Test set formed by flipping two pixels
Perfect classification for multiple runs on training set
Perfect classification on test set hardly possible (e.g. see pattern highlighted with red)
Classification performance (batch)
… test sets
znv
Batch training and …
z
v
n
13M. Prezioso et al., Nature May 2015M. Prezioso et al., IEDM’15 NanoXbar Workshop, July 2016
OVERCOMING NONLINEAR SWITCHING KINETICS
NanoXbar Workshop, July 2016 14
+log[V1(n)]
-log[δ1(n)]
0
0
-log[V1(n)]
0
0 +log[δ2(n)]
+log[V2(n)]
0
0
-log[V2(n)]
0
0-log[δ2(n)] +log[δ1(n)]
step 1 step 2
step 3 step 4
-2 -1 0 1 2 3
-1
-0.5
0
0.5
1
x 10-3
Voltage (V)
Current
(A
)
Experimental Data
Dynamic Model
Dynamic & Static Models
MODELING OF LARGE-SCALE CLASSIFIERS
I. Kataeva et al., IJCNN’15, M. Prezioso et al., IEDM’15
Main result: Comparable to the state-of-the-art classification performance for MNIST, GTSRB, and CIFAR benchmarks when using accurate models of hardware
Mis
cla
ssific
ation r
ate
Ratio of stuck devices (%)0 10 20 30 40 50 60 70 80 90 100
300-hidden-neuron MLP, variable-amplitude for in-situ training
Error rate is normalized with respect to the best software results (see table above)
2 1 0.4 0.2 0.1 0.04 0.02 0.01
100
10-1
10-2
10-3
3.0
2.5
2.0
1.5
1.0
0.5Norm
aliz
ed m
iscla
ssific
ation r
ate
Normalized STD for weight import (%)
~ 8 bit weight import precision
Dataset
Software Xbarin-situ (var. ampl.)
Xbarex-situ 2%
Xbarex-situ 0.2%
best average best average best average best averageMNIST 0.40 0.47±0.05 0.4 0.48±0.024 0.61 0.89±0.22 0.41 0.42±0.01GTSRB 1.36 1.53±0.18 1.26 1.56±0.27 1.42 1.56±01 1.46 1.47±0.01
CIFAR10 15.63 15.91±0.22 15.67 15.87±0.22 19.77 20.29±0.43 15.5 15.8±0.01
Experimentally-verified memristor device models
F. Merrikh Bayat et al., Applied Physics A, 2015
Classification performance results for large-scale deep learning convolutional neural networks
Performance sensitivity to defects and ex-situ training precision
15
NanoXbar Workshop, July 2016
16
CMOS/NANO HYBRIDS: THE IDEA
bottom nanowire level
top
nanowire level
similar
two-terminal
nanodevices
(“memristors”)
at each
crosspoint
WHAT:
• CMOS stack + simple nanoelectronic add-on
• nanowire / memristor crossbar
WHY:
• CMOS functionality and infrastructure intact
• potentially inexpensive fabrication
add-on
CMOSstack
NanoXbar Workshop, July 2016
17
DrainSource
Control Gate
Interpoly
DrainSource
Control Gate
DrainSource
Floating Gate
Control Gatedielectric
Gate
oxide
NVM (“FLASH”) MEMORY TECHNOLOGY
FG
CG
FG
CG
Channel
Channel Channel
NanoXbar Workshop, July 2016
18
NVM CELLS FOR ANALOG APPLICATIONS
J. Hasler and B. Marr (2013)
Example: “extended drain” NMOS structure
Hasler’s group at Georgia Tech (http://www-old.me.gatech.edu/mist/gokce.htm)
(from late 1990s: C. Mead, C. Diorio, P. Hasler,…)
NanoXbar Workshop, July 2016
NanoXbar Workshop, July 2016 19
bit line bit line
shared source line
word line 1 word line 2
common substrate
S
GG
D D
SILICON STORAGE TECHNOLOGY, INC. (SST): ESF1
0 0.2 0.4 0.6 0.8 10
20
40
60
|VDS
(V)|
|ID
S(
A)|
Fully Erased, VS=0, V
D=swept
Fully Erased, VD=0, V
S=swept
0 1 2 3 4 51E-13
1E-11
1E-9
1E-7
1E-5
VG
(V)|I
DS(A
)|
Fully Erased,VDS
=1V
Fully Erased,VDS
=-1V
Fully Programmed,VDS
=1V
Fully Programmed,VDS
=-1V
(a)
(b) VG=4V
3.5V
3.0V
2.5V
2.0V
1.5V
1.0V
Output current as a function of applied voltages:
F. Merrikh-Bayat et al. (2015)NanoXbar Workshop, July 2016
20
FLASH ARRAY REDESIGN FOR ANALOG APPLICATIONS
VSE = 0V
half-selected cell (type D)
3.8
4μ
m
1.36μm
3.8
4μ
m
3.12μm
VGE
0V
VDE0V 0V
VSE’
0V
VGE = 11V
0V
0V
0V
0V
VDE=0V VD
E ’
half-selected cell (type C)
selected cell
VDP=0V
VSP = 7.6V
0V
VDP ’ > 2V
VGP = 1.6V
0V
0V
selected cell half-selected cell (type A)
half-selected cells (type B)
VGP
VSP
0VVGP’
0V
0V
selected cell selected cellhalf-selected cells (type B)
half-selected cell (type E)
VGP’ = 0V
VGP’
Old array:
New array:
“program” “erase”
“program” “erase”
F. Merrikh-Bayat et al. (2015)<100F2 area per synapse
NanoXbar Workshop, July 2016
21
0 20 40 60 80 1000.95
1
1.05
Read Pulse Number
Cu
rre
nt
(nA
)
0 20 40 60 80 1000.99
0.995
1
1.005
1.01
Read Pulse Number
Cu
rre
nt
( A
)
0 100 200 300 400 500 600 700 800 900
1E-9
1E-8
1E-7
1E-6
Pulse Number
Cu
rre
nt
(A)
TUNING (OF EACH CELL!) TO PRE-SET VALUES
F. Merrikh-Bayat et al. (2015)NanoXbar Workshop, July 2016
22
-1 -0.5 0 0.5 1 -0.1
-0.05
0
0.05
0.1
inI
inI
outI
outI
2ini
2ini
Vdd
w1+
w1-
w1-
w1+
ddV
Ibias = 500 nA
Vdd = 1V
w1+ ≈ 10-5
Iout+ - Iout
- = (w1+ - w1
-)(Iin+ - Iin
- )
I ou
t+-
I ou
t-(µ
A)
Iin+ - Iin
- (µA)
-1 -0.5 0 0.5 1-0.115
-0.11
-0.105
-0.1
-0.095
-0.09
-0.085
Iin
+-I
in
- (A)
Slo
pe
VECTOR-BY-MATRIX MULTIPLIER (VMM) DEMO
VMM linearity:
F. Merrikh-Bayat et al. (2015)
Transfer to a better NVM 2D technology would bring
synapse area well below ~ 1 μm2.
NanoXbar Workshop, July 2016
SPIKING NEURAL NETWORKS
23
-60 -40 -20 0 20 40 60-100
-75
-50
-25
0
25
50
75
100
W
(
)
t(ms)
-60 -40 -20 0 20 40 60
-1.0
-0.5
0.0
0.5THRESHOLD
Min
Max
Vo
lta
ge
(V
)
t (ms)
RESET
THRESHOLD
SET
0 20 40 60
-0.6
-0.4
-0.2
0.0
0.2
0.4
0.6
Vo
lta
ge
(V
)
t (ms)
Pre
Post
- Richer functionality (spatial
and temporal processing) and
better energy efficiency of
spiking networks as compared
to firing rate
- Local (Hebbian) training
more efficient hardware
- Essential feature to
demonstrate: Spike-timing
dependent plasticity (STDP)
Motivation Experimental demonstration of STDP
- Three STDP windows
demonstrated using crossbar
- The most accurate STDP
demonstration to date
no
n128in
n2in
n1in
𝐺0
𝐺1
𝐺128
𝑆1in(𝑡)
𝑆2in(𝑡)
𝑆128in (𝑡)
nd
𝑆o(𝑡)
𝑆d(𝑡)
teacher
output leaky
integrate-and-
fire neuron
input
neurons
synapses
time
G𝑆d(𝑡)𝑆in(𝑡)
𝑡d − 𝑡in
time
G𝑆𝑜(𝑡)𝑆in(𝑡)
𝑊o
𝑡o − 𝑡in
0 50 100 150 200
Time (ms)
0 5 10 15 20 25 30
128
96
64
32
15
10
5
01.5
1.0
0.5
0
2015 10 50
Epoch #Pe
rfo
rma
nce
index
Neuro
n
pote
ntialE
poch #
Neuro
n #
𝑆in
𝑆d
𝑆o
threshold
STD/mean
10
5
0
<P
erf
orm
ance i
ndex>
CombinedDevice-to-deviceCycle-to-cycle Initial state
no spiking
1000 training
epochs
STD/mean
<P
erf
orm
ance i
ndex>
50100 -50 0 50
-50
0
50
100
150
200
t (ms)W (S)
W
(%
)Δ
G(%
)
Δt (ms)
2040
6080
100
-50
-25
0
25
50
75
100
125
150
175
200
-60 -40 -20 0 20 40 60
100 S
75 S
50 S
25 S
W
(%
)
t (ms)
Weight (S)
ΔG
(%)
,GtG
,0,1tanh
,0,1tanh
tctba
tctba
tt
tt
t
.0 ,tanh1
,0,tanh1
0
0
tcGb
tcGb
GG
GG
G
Experimentally-verified analytical model STDP
Simulation of memristor-based spiking neural networks
M.Prezioso et al., Nature Scientific Report, 2016
M. Prezioso et al., ISCAS 2016NanoXbar Workshop, July 2016
24
NanoXbar Workshop, July 2016
SUMMARY
Emerging nonvolatile memories enable (for the first time?) efficient analog neural network implementations and could challenge human brain in energy efficiency and speed Experimental demonstration of key hardware block for both memristor
and flash-based artificial neural networks Small scale demonstrations of firing-rate feedforward/recurrent and
spiking memristor-based artificial neural networks with comparable to state-of-the-art functional performance for large scale NVM-based networks via simulation with data-verified device models
Estimated >100x / >1000x improvement in energy efficiency as compared to ASICs for flash / memristor based implementations
Need industry involvement to develop large-scale memristor circuit no such issue with flash memory–based circuit
Strukov et al., DRC’16
SELECTED RECENT PUBLICATIONS
26
NanoXbar Workshop, July 2016
• F. Merrikh-Bayat, X. Guo, M. Klachko, N. Do, K. Likharev, and D. Strukov, “Model-based high-precision tuning of NOR flash memory cells for analog computing applications”, to appear in Device Research Conference (DRC’16), Newark, DE, June 2016 (NOR flash)
• M. Prezioso, Y. Zhong, D. Gavrilov, F. Merrikh Bayat, B. Hoskins, G. Adam, K.K. Likharev, and D.B. Strukov, “Spiking Neuromorphic Networks with Metal-Oxide Memristors”, to appear in International Symposium on Circuits and Systems (ISCAS'16), Montreal, Canada, May 2016 (Memristor spiking neural networks)
• M. Prezioso, F. Merrikh Bayat, B. Hoskins, K. Likharev, and D. Strukov, "Self-adaptive spike-time-dependent plasticity of metal-oxide memristors", Nature Scientific Reports 6, art. 21331, Jan. 2016. (Memristor spiking neural networks)
• F. Merrikh Bayat, M. Prezioso, X. Guo, B. Hoskins, D.B. Strukov, and K.K. Likharev, "Memory technologies for neural networks", in: Proc. IMW'15, Monterey, CA, May 2015, pp. 1-4. (brief review)
• F. Merrikh Bayat, X. Guo, H.A. Om'mani, N. Do, K.K. Likharev, and D.B. Strukov, "Redesigning commercial floating-gate memory for analog computing applications", in: Proc. ISCAS'15, Lisbon, Portugal, May 2015, pp. 1921-1924. (NOR flash)
• M. Prezioso, F. Merrikh Bayat, B.D. Hoskins, G.C. Adam, K.K. Likharev, and D.B. Strukov, "Training and operation of an integrated neuromorphic network based on metal-oxide memristors", Nature 521, pp. 61-64, May 2015. (Memristor firing-rate MLP networks)
• X. Guo, F. Merrikh-Bayat, L. Gao, B. D. Hoskins, F. Alibart, B. Linares-Barranco, L. Theogarajan, C. Teuscher, and D.B. Strukov, "Modeling and experimental demonstration of a Hopfield network analog-to-digital converter with hybrid CMOS/memristor circuits", Frontiers in Neuroscience 9, art. 488, Dec. 2015. (Memristor recurrent networks)
• F. Merrikh Bayat, B. Hoskins, and D.B. Strukov, "Phenomenological modeling of memristive devices", Applied Physics A 118 (3), pp. 770-786, 2015. (Memristor model)
• M. Prezioso, I. Kataeva, F. Merrikh-Bayat, B. Hoskins, G. Adam, T. Sota, K. Likharev, and D. Strukov, "Modeling and implementation of firing-rate neuromorphic-network classifiers with bilayer Pt/Al2O3/TiO2-x/Pt memristors", IEDM'15, Dec. 2015. (Memristor firing-rate MLP networks)
• M. Payvand, A. Madhavan, M. Lastras-Montaño, A. Ghofrani, J. Rofeh, K.-T. Cheng, D. Strukov, L. Theogarajan, "A configurable CMOS memory platform for 3D-integrated memristors", in: Proc. ISCAS'15, Lisbon, Portugal, May 2015, pp. 1378-1381. (Memristor integration)
• I. Kataeva, F. Merrikh Bayat, E. Zamanidoost, and D.B. Strukov, "Efficient training algorithms for neural networks based on memristive crossbar circuits", in: Proc. IJCNN'15, Killarney, Ireland, July 2015, pp. 1-8. (Memristor firing-rate MLP networks modeling)
• F. Alibart, E. Zamanidoost, and D.B. Strukov, "Pattern classification by memristive crossbar circuits with ex-situ and in-situ training", Nature Communications 4, art. 2072, 2013 (Memristor firing-rate MLP networks)
• J.J. Yang, D.B. Strukov and D.R. Stewart, “Memristive devices for computing”, Nature Nanotechnology 8, pp. 13-24, 2013 (review)