ESA, ACT ESTEC April 29, 2015
Memristive Tunnel Junctions for Neuromorphic Circuits Andy Thomas Fakultät für Physik
Acknowledgments
L. Schnatmann
N. Shepheard
M. Schirmer
J. Sterz
O. Simon
Center for Spinelectronic Materials and Devices S. Niehörster
Z. Kugler
I.-M. Imort
J. Münchenberger
S. Fabretti
O. Schebaum
M. Schäfers
G. Reiss
Ministerium für Innovation, Wissenschaft und Forschung des Landes NordrheinWestfalen
A. Hütten
2
Collaborations Spin electronic C. Felser, MPI Dresden J. Moodera, M.I.T. T. Kampfrath, FHI Berlin K. Nielsch, U Hamburg
Spin caloritronics
Memristive systems C. Kaltschmidt, U Bielefeld E. Chicca, U. Rückert, CITEC Bielefeld
memristor
M. Münzenberg, U Greifswald S. Goennenwein, WMI Garching C. Heiliger, U Gießen M. Kläui, U Mainz
3
Why neuromorphic circuits? 1 Introduction Scientific curiosity Maximum reduction?
R Lochkartenleser
Avoid von Neumann bottleneck
Memristors have attracted great interest for a variety of applications in recent years. CC An obvious use would be as a memory device [17, 52, 50] or, more ambitiously, a reconfigurable logic device [88, 10, 89, 11, 64]. However, the most interesting Kontrolleinheit implementation of memristivepostsynaptic devices is neuromorphic computing. cell im Neuromorphic pu computing aims to use biological mechanisms operating within lse s synapse the brain as a blueprint to construct novel computer architectures. Carver Mead built the foundation of this field and proposed large-scale adaptive analogue systems axon presynaptic cell because of their robustness as well as good power efficiency [61]. The efficiency of Arithmetische Speicher these systems is particularly promising, as shown in Table 1. Einheit synapse
dendrite
CA
M
Table 1 Comparison of the power consumption of three different technologies [74]. A biological neuron draws less power and consumes less area than a digital computer or silicon neuron.
Energy consumption (J/spike) Size (µm2 )
Digital computer Silicon neuron
Biological neuron
10 5 108
10 10
[74] C.S. Poon, Frontiers in neuroscience 5 (2011) 1
10 8 3 ⇥ 103
11
Efficiency
Biological neural networks
Biological neural networks need neurons and synapses
neurons are connected via the synapses.
postsynaptic cell
im pu
lse
s
synapse
Neurons integrate signals and fire when exceeding threshold
axon
presynaptic cell
dendrite
synapse
voltage threshold triggers pulse I(t)
im
pu
C
lse
s
presynaptic cell
R
pulse triggers switch
postsynaptic cell
leaky integrate and fire
synapse voltage threshold triggers pulse I(t)
Thomas, J. Phys D: Appl. Phys. 46 (2013) 093001, Thomas, Kaltschmidt, Memristor Networks, Ed. Adamatzky, Chua (2014) 151-172
oversampled ΣΔ modulators
= 1-Bit AD converter 6
Biological synapse Axon
Pre-synpase
Glu
synaptic cleft
vesicle fusion
Glu
Glu Glu
Glu
vesicles containing neurotransmitters
Glu
Post-synapse Ca2+
gene expression via CREB, NF-kappaB
Simplify the mechanisms via a simple model: One effective connection strength. Mayford, Siegelbaum, Kandel, Cold Spring Harbor Perspectives in Biology 4(6), a005,751 (2012)
7
Electronic symbols J. Phys. D: Appl. Phys. 46 (2013) 093001
Topical Review
voltage threshold triggers pulse I(t)
im pu
synapse
lse
s
presynaptic cell C R
excitatory
postsynaptic cell pulse triggers switch
inhibitatory
neuron
Figure 13. Schematic symbols for synapses and neurons as used in the following paragraphs.
Symbols
of the research on this topic and explained the reasoning in the following way [75]: For many problems, particularly those in which the input data are ill-conditioned and the computation can be specified in a relative manner, biological solutions are many orders of magnitude more effective than those we have Figure 12. Integrate-and-fire circuits: the perfect integrate-and-fire model consists of a capacitor, threshold detector and switch (without been able to implement using digital methods. [...] LargeReview 46resistor). (2013) 093001 Topical Review Once the voltage reaches a threshold,Topical the spike is fired and scale adaptive analog systems are more robust to component the switch is closed to shunt the capacitor. In the leaky version, a 093001 Review systems, degradation and failure than are moreTopical conventional resistor is added that slowlyvoltage drains the capacitor with time. This threshold excitatory excitatory and they use far less power. The need for less power is corresponds to leakage current through a membrane in a living cell. triggers pulse voltage threshold particularly obvious if we compare the performance of the excitatory triggers pulse brain of even an invertebrate with a computer CPU and contrast integrate-and-fire [68–71]. The rather simple model captures the power consumption. However, there are some tasks that neuron setwo of the most important aspects of neuron neuronal excitability;synapse are difficult for a human and easy for a computer, such as inhibitatory inhibitatory specifically, the neuron integrates the incoming signals and multiplying two long numbers, and synapse neuron other problems that a inhibitatory generates the spikes once a certain threshold is Figure human cansymbols easily solve computers fail to solve. 13. Schematic for but synapses and neurons as used in ematic symbols for synapses and neurons as usedexceeded in (figure R 12). In 2008, the use of memristors to mimic biological the following paragraphs. aragraphs. Figure 13. Schematic symbols synapses and neurons as used in pulse triggersusing two simple electric mechanisms,forsuch This behaviour is often explained as STDP, was already hypothesized the following paragraphs. switch circuits [72]. pulse The perfect [76]. Snider implemented the spike time dependence using triggersintegrate-and-fire circuitofconsists the research on this topic and explained the reasoning in the capacitor, threshold detectorthe andreasoning switch. Once a spike memristive h of onathis topicswitch and explained in the nanodevices as synapses and conventional CMOS 8 following is fired, the switch closes and shunts the capacitor. An way [75]: For many problems, particularly those in synapse
presynaptic cell
postsynaptic cell
Biological synapses 100
potentiation (%)
Exhibit LTP, LTD, STDP
Long-term potentiation (LTP)
50
3
input
γ
0
2 -50
β
1 0
1
2 time (h)
3
α
4
change (%) synaptic strength
outp
(a)
100
1.0
Long-term depression (LTD)
ut
Spike-time dependent plasticity (STDP)
(b)
50
0.5
0 -50
0.0
-60
40
60
T. Bliss at al, Nature 361 (1993) 31, Y. Goda et al, Neuron 16 (1996) 103, S. Cassenaer et al, Nature 448 (2007) 709
9
-10
0
10
20
30
-40
-20 0 20 spike timing (ms)
time (min)
Observations in a biological neural network 7 The synapses exhibit LTP (cooperative, associative, and input-specific), LTD, and STDP.
eported that ltp is characterized by three basic ity and input-specificity [43]. We will discuss these g paragraphs. CONTENTS
Cooperative
8
Associative
G. Barrioneuvo, Natl. Acad. Sci. 80 (1983) 7347
Figure 8. The application of the weak stimulus W or a strong stimulus S does not lead to a potentiation. If W and S are applied at the same time, long-term potentiation can be observed. Reprinted from Proc. Nat. Acad. Sci. [48], Copyright (1983).
Input-specific: post tetanus (min) required Pre by 5 general 10 15 B.L. McNaughton, Brain. Res. 157 (1978) 277
d high-intensity stimulus (b) was applied to a fibre. The lowced a reaction, but it was short lasting. The high-intensity
causality
Tetanised pathway (%) 100
390 380 332
Spike timing dependent plasticity (STDP) 1
3
input
γ 2
β
1
Spike-time dependent plasticity (STDP)
2 output
α outp
(a)
ut
coincidence detector
change (%)
α 100
potentiate α
(b)
1
50
2
0 -50 -60
-40
-20 0 20 spike timing (ms)
40
60
output depress α
S. Cassenaer et al, Nature 448 (2007) 709
11
Memristor
Thin film samples @ Bielefeld Thin film devices Memristors: artificial neural networks integration with CMOS hybrid samples with biological neurons
Materials ‘development’: Ferromagnets, e.g. Heusler (HMF) SC: e.g. Heusler, MgB2 Oxides: e.g. Ta-O, HfO2, MgO Insulator Metal Lithography: optical e-beam ion beam etching Spintronic devices: sensors memory logic
Spin caloritronics: TMS vs TMR thermal spin transfer torque spin pumping
13
Sample preparation Thin film (co-)deposition on Si-wafer, MgO, STO, ..., substrates Optional in- and ex-situ anneal, ex-situ optional field cool Lithography (e-beam, optical), subsequent ion-beam etching Transport measurements (0.3-500K, up to 4T)
Magnetic Tunnel Junction
Meservey-TedrowTunnel Junction
Memristor
Ferromagnet
Superconductor
Metal
Insulator
Insulator
Insulator
Ferromagnet
Ferromagnet
Metal
Examples of the prepared thin film devices
14
Memristive tunnel junctions
!"$*+()$ !"#$%&'()$ ,-$*+()$
Electrically controlled, two terminal, bipolar, analog devices
Metal
10 5 0 -5 -10
Current [mA]
current
Abbildung 1: Stack
0
Insulator Metal
v=v0 sin(ωt)
Ta/TaO/Pd
-0.4
0 voltage
0.0 Voltage [V]
0.4
152 Abbildung 2: Loop mit Switching Effekt von 120% (MgO 12%)
resistance (Ω)
• • • ••
151
•
•
•
Die oberen 3 Kurven wurden aus der gleichen Probe generiert. Die Barriere wurde hergestellt indem 2nm Ta mit Sauerstoffplasma oxidiert wurden. Die Beschleunigugnsspannung betrug -80V und die Oxidationszeit betrug 200s.
•
150
• •
• • • • •
• •
149 0
20
40 flux (Vs)
•
•
• •
60
Thomas, J. Phys D: Appl. Phys. 46 (2013) 093001, Krzysteczko,..., AT, Appl. Phys. Lett. 95 (2009) 112508 Semicond. Sci. Technol. 29 (2014), guest Editor: A. Thomas
80
current flows one direction => turn knob one direction
current flows other direction => turn knob other direction
+I: resistance up
-I: resistance down 1
15
Memristor based synapses
(V)
mains stable when the activity is terminated.
Long term potentiation and depression 2 4 6 8 10 12
-0.5 0
Time (min)
2
Biological System 100
223 3
2
221
1
R 220 (Ω) 219
-50 0
218
1
2 time (h)
3
4
neuron
Successive pulses cause successive potentiation/ depression
0
0
1.0
0.5
Figure 50: The effect of 0.0 subsequent train application. -10 0 10 20 The resistive states are numtime (min) bered. The state zero is the result of the refreshing procedure shown in fig.at51. T. Bliss al,The Nature 361 (1993) other states are induced by 31, Y. Goda et al, Neuron 16 standard spike trains.
4
222
50
neuron
synaptic strength
time (min)
potentiation (%)
conductance change (%)
LTP in figure 1 48 are applied. They are characterized by only two parameters: the amplitude of the sine profile number I vmax and theexcitatory of spikes per train Nsp . 0 A typical measurement is shown in figure 49. The graph consists 0.5 of three traces. The bottom trace plots the applied voltage. synapse inhibitatory I The middle trace displays the corresponding flux. The top trace excitatory shows that due to the voltage treatment, the resistance is driven 0.0 from a lower stable level to a higher stable level. This is exactly how an artificial synapse should respond.LTDThe activity at synapse inhibitatory -0.5 the synaptic connection leads a stable modification thereof—the activity is0remembered. 5 10
30
(1996) 103
217 0
10
20
30
Time (min)
40
50 17
Pulse shaping for spike timing dependent plasticity 3
input
γ 2
β
1
Take real properties
into account
R 184 (Ω) 180
α outp
-0.55
ut
change (%) (b)
RRL,P L
176.0
(a)
100
RRH H,P
190.9 188
vpre
t>0
vpre
50
0.0 Voltage (V)
0.55
t1000 Neurons) Neuromorphic nanoscale memristor synapses Vdd Vdd
Vdd
I
Vdd
Iw
Ith
Vdd
Vw
IwN
Vw Vw Vw
VinN
Vw
Vin4
Isyn
existing neuromorphic chip design + memristor pads + additional electronics = neuronal circuit with synaptic weights
Vin3 Vin2 Vin1
(a)
(b)
chip design (AG E.synapse. Chicca, Citec Bielefeld), circuit implementing a Figure Neuromorphic 6: Neuromorphic memristive (a) Schematic array of memristive synapses, with independent inputs and synaptic weights, but wi Memristor preparation and lithography (AG A. Thomas, U Bielefeld) shared temporal dynamics. (b) SPICE simulations of the circuit in Fig. 6a showing t output Isyn EPSC in response to a pre-synaptic input spike, for 4 di↵erent memrist G. Indiveri et al., Nanotech. 24 (2013) 384010 23 conductance values.
Take home message Memristive System potentiation (%)
100
LTP
1 I
excitatory
0 synapse
inhibitatory
0.5 I
0.0
excitatory
LTD
synapse
inhibitatory
-0.5
neuron
50
0
-50
neuron
0
synaptic strength
conductance change (%)
2
Biological System
1.0
5
2 0.5 1
α (a)
Δ R (Ω)
STDP
-10
0
10
outp
20
ut
30
time (min)
(b)
50 0 -50
0
-60
-40
-20 0 20 spike timing (ms)
40
60
-4 -8 -200
450
(Ω)
Change (%)
100
4
4
β
change (%)
8
3
γ
10 time (min)
2 time (h)
3
input
0.0
0
1
back-hopping
-100
0 spike timing (s)
100
200
T. Bliss at al, Nature 361 (1993) 31,
Y. Goda et al, Neuron 16 (1996) 103
S. Cassenaer et al, Nature 448 (2007) 709
25