Académique Documents
Professionnel Documents
Culture Documents
Early attempts
The idea of making custom analog VLSI implementations of neural networks dates back to the late 80s - early 90s:
[Holler et al. 1989, Satyanarayana et al. 1992, Hammerstrom 1993, Vittoz 1996]
General purpose computing Full-custom analog implementation Neural network accelerator PC-boards
Current research
Technological progress Power-dissipation/computational power Application-specic focus Embedded system integration
3 / 78
ICANN09 Tutorial
1 / 78
ICANN09 Tutorial
Above threshold (strong inversion) Mixed analog/digital Rate-based Real-time Conductance-based Large-scale, event-based networks
Below threshold (weak inversion) Fully analog Spiking Accelerated-time Integrate-and-Fire Small-scale, hard-wired
ICANN09 Tutorial
4 / 78
ICANN09 Tutorial
5 / 78
Most designs can be traced back to one of two types of silicon neurons,
MOSFETs in subthreshold
Vd Vg Ids Vs
where I0 denotes the nFET current-scaling parameter n-FET subthreshold transfer function Ids = I0 en Vg /UT eVs /UT eVd /UT
Vg Vs Ir Vs Ir
Qs
Vd If Vg Vd If
Qd
Ids = I0 en Vg /UT eVs /UT eVd /UT is equivalent to: Ids = Ids = I0 e
V Ug U s T V T
10 10 10
-2
-3
subthreshold
-4
I0 e Ir
U g Ud
T
10 10
-5
T
-6
Ids (A)
If
10 10 10 10 10 10
-7
above threshold
If Vds > 4UT the Ir term becomes negligible, and the transistor is said to operate in the saturation regime: Ids = I0 en Vg /UT Vs /UT
-8
-9
-10
-11
-12
0.5
1.5
2.5
3.5
4.5
Vgs (V)
ICANN09 Tutorial
10 / 78
ICANN09 Tutorial
11 / 78
Vd
Vs
Vg Vs
Vg Vd
Vb
Vin
I out
M1 M2
Vd
Inverting amplier
Differential pair
I1 V1 I2 V2
In traditional CMOS circuits, all n-FETs have the common bulk potential (Vb ) connected to Ground (Gnd), and all p-FETs have a common bulk potential (typically) connected to the power supply rail (Vdd ). The corresponding (complementary) equation for the p-FET is Ids = I0 e
G.Indiveri (NCS @ INI)
Vin
Vout
Vbn
M2
Vs M1
M3
p (Vdd Vg )/UT
(Vdd Vs )/UT
(Vdd Vd )/UT
ICANN09 Tutorial
12 / 78
ICANN09 Tutorial
14 / 78
The differential-pair
V1 Vs
UT
I1 = I0 e I2 = I0 e
Vdd
M4
Vdd
M5
V1
V2 Vs
UT
M2
Vs M1
M3
Ib = I1 + I2 = I0 e UT e
V Us T
Vb
Vbn
Iout I1 I2 Vs Ib Vb
M3 M2
Vout V2
Iout = Ib tanh
2UT
(V1 V2 )
Ib I0
V1
1 e UT + e UT
V1 V2
1 0.8
x 10
V1
I1
M1
In the linear region (|V1 V2 | < 200mV ): Iout gm (V1 V2 ) where gm = Ib 2UT
I2
I1 = Ib e I2 = Ib
e UT
V1
UT
+e
V1
V2
UT
I1, I2 (A)
0.6
0.4
e UT e UT + e UT
V1 V2
0.2
0 0.3
0.2
0.1
0 V1V2 (V)
0.1
0.2
0.3
is a tunable conductance.
ICANN09 Tutorial
15 / 78
ICANN09 Tutorial
16 / 78
What is a synapse?
In 1897 Charles Sherrington introduced the term synapse to describe the specialized structure at the zone of contact between neurons as the point in which one neuron communicates with another.
Electrical | Chemical
Vmem Gl GABA Einh (K+, Cl, ...) Cmem
2005 winner of the Science and Engineering Visualization Challenge. by G. Johnson, Medical Media, Boulder, CO.
ICANN09 Tutorial
19 / 78
ICANN09 Tutorial
20 / 78
Synaptic transmission
In chemical synapses the presynaptic and postsynaptic membranes are spearated by extracellular space. The arrival of a presynaptic action potential triggers the release of neurotransmitter in the extracellular space. The neurotransmitters react with the postsynaptic receptors and depolarize the cell. Chemical synaptic transmission is characterized by specic temporal dynamics.
FIG .
maximal preganglionic stimulation. Superimposed Superimposed excitatory EPSCs recorded in a sympathetic neuron at different membrane potentials between 010 and NATURE OF THE POSTSYNAPTIC RECEPTOR ACTIVATED BY 0100 mV (in 10 mV steps) under 2-electrode voltage-clamp conditions. post-synaptic currents (EPSCs) s intervals, to NATURAL AC . To correctly model and reproduce evoked Excitatory post-synaptic potential Cell was held at 050 mV and then repeatedly stepped, at 20 it was important whether or membrane levels at observed. Note presence recorded 10/ 0which synaptic current wasCa concentration was 5 synaptic currents,response toto determinethe nicotinic (EPSP) in multiple 30 neuron at different of fast I in 0 in a mV tracings. External not conductances other than those associated to mM. receptor were activated by nerve-released ACh. In fact, ACh membrane potentials (from Sacchi et pre-synaptic spikes terminals is thought to genrelease by the presynaptic nerve(from Nicholls et al. cell was estimated in culture, broad series addition al., 1998). for the rat sympathetic neuron putative M- erate in the ganglion asustaining theof side effects, intransmis1992). whereas the activation-deactivation time constants for to the central role of fast excitatory
H
2/ Na
2.
low the linear t. This might be partly because of inward rectication by the ACh-evoked current (Fieber and Adams 1991; Mathie et al. 1990), but it is at least partly accounted for by the changes in ionic gradients that are associated with even small ionic ows (especially accumulation of K / ions in the perineuronal space) (Belluzzi and Sacchi 1990). In all neurons the decay of EPSC was well t by a single exponential function. The average time constant, t, was 7.5 ms at 075 mV (n 14) and was scarcely voltage-dependent: it decreased exponentially with a constant of 260.4 mV from 0105 to 025 mV (Fig. 3A). EPSC amplitudes decreased in the average (5 neurons) to 69% of the initial value when Ca 2/ concentration in the bath was lowered from the usual 5 to 2 mM; the current decay time constants were also somewhat decreased (by 21% in the average). The voltage dependence of the decay time constant and the estimated reversal potential did not appear to be affected by calcium concentration (Fig. 3B).
ICANN09 Tutorial
21 / 78
sion. Autoreceptors were proposed to control evoked transmitter release (see Starke et al. 1989, for a review). The evidence for a physiological role in the ganglion seems to R E S U L G.Indiveri (NCS @ INI) TS ICANN09 be greatest for the muscarinic autoreceptors, which are pos-/ 78 Tutorial 22 tulated to mediate a negative feedback loop that reduces Synaptic current at the sympathetic ganglionic neuron transmitter mobilization (Koelle 1961; Koketsu and Yamada
currents were 220 ms in the 060/ 030 mV voltage range (Owen et al. 1990).
In classical neural network theory signals are (tipically) continuous values that represent the neurons mean ring rate, neurons implement a saturating non-linearity transfer function (S) on the inputs weighted sum, the synapse implements a multiplication between the neurons input signal (Xi ) and its corresponding synaptic weight (wi ).
G.Indiveri (NCS @ INI) ICANN09 Tutorial 24 / 78
Ib
Figure: Schematic of half of a Gilbert multiplier. This circuit multiplies Iin1 and Iin2 by Ib if Iin1 + Iin2 = Ib .
G.Indiveri (NCS @ INI) ICANN09 Tutorial 26 / 78
Wi
Vi =
IWi Cmem
In pulse-based neural networks the weighted contribution of a synapse can be implemented using a single transistor. In this case p-FETs implement excitatory synapse, and n-FETs implement inhibitory synapses. The synaptic weight can be set by changing the Wi bias voltage or the t duration.
Vg
Id
Vg(t)
Vdd
Vc C
Id =
d C dt Vc
Id (t ) = I0 e UT
Id(t)
(Vdd Vg (t ))
ICANN09 Tutorial
27 / 78
ICANN09 Tutorial
28 / 78
Iw = I0 e
V M I Vw Csyn Vsyn Msyn Isyn
, c , d
(Vsyn Vw )
UT
I
Csyn UT
I = I0 e Ic = C
(Vdd V )
UT
I = I0 e Ic = C
(Vdd V )
d dt
d dt
(Vdd Vsyn )
(Vdd Vsyn )
UT
Mw Iw Mpre
(Vdd Vsyn )
(Vdd Vsyn )
UT
Isyn = I0 e d dt
Isyn = I0 e Isyn (t ) =
+ (t ti ) I e c
syn
+ + (t ti ) d I e
d
UT dt
Vsyn
syn
Isyn Iw
(Vw Vdd )
UT
Isyn (t ) = I0 e
t c f c(c +d ) t d
, with f =
n t
, f<
Isyn + Isyn =
, Iw0 = I0 e
ICANN09 Tutorial
29 / 78
ICANN09 Tutorial
30 / 78
DPI equations
DPI transfer function: d Iout + Iout dt 1 Iout
Iout Ig
M I Iin
Csyn Vsyn
I = I0 e Ic = C
(Vdd V )
UT
I 1 + Ig I Iin ,
Iin
Vthr
d dt
Mthr
Min
(Vdd Vsyn )
Vsyn
d dt
Iout + Iout
if Iout
Ig , Iw
Vw
Mw Iw Mpre
Iin = Iw e
e
Vsyn
UT
UT
+e
UT
Iout (ti +t ) =
Ig Iin I
1 e
ti ti 1
(Vdd Vsyn )
Isyn =
UT
Isyn
e ( )d
Iw Igain I
(Bartolozzi, Indiveri, 2007)
t , =
t + ISI
32 / 78
ICANN09 Tutorial
31 / 78
ICANN09 Tutorial
300 Vw=420mV 250 200 150 100 50 0 0 Vw=440mV Vw=460mV EPSC (nA)
450 400 350 300 250 200 150 100 50 0.05 0.1 Time (s) 0.15 0 0 0.5 1 1.5 2 2.5 Time (s) 3 3.5 Vw=320mV Vw=340mV
Vw=300mV
100 80 60 40 20 0
EPSC (nA)
50
200
ICANN09 Tutorial
33 / 78
34 / 78
Short-term depression
Vgain Va M6 M7 Id Vx M5 C Vpre Isyn M4 C2
Vx (V) 0.35
Slow recovery
l e e f e h
Va
Vw Cd
Vd
M1
Ir
Vd
Ir
M2
0.15
e n
Vpre
M3
0.1 0.04
0.06
0.08
0.14
0.16
(a)
(b)
ICANN09 Tutorial
35 / 78
P,
p qYdi gh$
uS t
t S r sS @ S i Ygh$ f
Figure 1: Schematic for a depressing synapse circuit and responses to a regular input spike (M. Boegerhausen, P. Suter, and S.-C. Liu 2003) train. (a) Depressing synapse circuit. The voltage determines the synaptic conductance while the synaptic term or is exponential in the voltage, . The subcircuit consisting of transistors, , , and ICANN09 Tutorial the dynamics of , control . The presynaptic G.Indiveri (NCS @ INI) 36 / 78 input goes to the gate terminal of which acts like a switch. When there is a presynaptic
e"d
A VLSI recurrent network of integrateandre neurons connected by plastic synapses with long term memory. IEEE Transactions on Neural Networks, 14(5):12971307, September 2003. P. Higer, M. Mahowald, and L. Watts. A spike based learning neuron in analog VLSI. In M. C. Mozer, M. I. Jordan, and T. Petsche, editors, Advances in neuralinformation processing systems, volume 9, pages 692698. MIT Press, 1997. G. Indiveri, E. Chicca, and R. Douglas. A VLSI array of low-power spiking neurons and bistable synapses with spiketiming dependent plasticity. IEEE Transactions on Neural Networks, 17(1):211221, Jan 2006. S. Mitra, G. Indiveri, and S. Fusi. Learning to classify complex patterns using a VLSI network of spiking neurons. In J.C. Platt, D. Koller, Y. Singer, and S. Roweis, editors, Advances in Neural Information Processing Systems 20, pages 10091016, Cambridge (MA), 2008. MIT Press.
bistability: use two synaptic states; redundancy: implement many synapses that see the same pre- and post-synaptic activity stochasticity & inhomogeneity: induce LTP/LTD only in a subset of stimulated synapses.
- Slow learning: only a fraction of the synapses memorize the pattern. + The theory is matched to the technology: use binary states, exploit mismatch and introduce fault tolerance by design.
G.Indiveri (NCS @ INI) ICANN09 Tutorial 37 / 78
ICANN09 Tutorial
38 / 78
Neurons . . . in a nutshell
A quick tutorial
Real Neurons Complexity Conductance based models Integrate and re models Rate based models
Sigmoidal units Linear threshold units
(adapted from B. Mel, 1994) G.Indiveri (NCS @ INI) ICANN09 Tutorial 40 / 78 G.Indiveri (NCS @ INI) ICANN09 Tutorial 42 / 78
Equivalent Circuit
Eex (Na+, ...) Glutammate
Vmem gK EK
If the membrane voltage increases above a certain threshold, a spike-generating mechanism is activated and an action potential is initiated.
G.Indiveri (NCS @ INI) ICANN09 Tutorial 44 / 78
Gl
Cmem
If excitatory input currents are relatively small, the neuron behaves exactly like a rst order low-pass lter.
ICANN09 Tutorial
43 / 78
Spike properties
Iin=I1
Refractory Period
Pulse Width
Iin=I2 > I1
Refractory Period
The rst articial neuron model was proposed in the 1943 by McCulloch and Pitts. Hardware implementations of this model date almost back to the same period.
One of the most inuential circuits that implements an integrate and re (I&F) model of a neuron was the Axon-Hillock Circuit, proposed by Carver Mead in the late 1980s.
In 1991 Misha Mahowald and Rodney Douglas proposed a conductance-based silicon neuron and showed that it had properties remarkably similar to those of real cortical neurons.
G.Indiveri (NCS @ INI) ICANN09 Tutorial 49 / 78
ICANN09 Tutorial
47 / 78
Sodium
1V Vm
+
VNa Vthr
+
GNaoff
INaoff
ENa
[Ca]
Vthr Vmem
INaon
INa
+
GNaon
Passive Leak
Vdd Eleak Vmem
Vm [Ca]
Vdd
+
Gleak Cmem
+
VK Vthr
Passive
IK GK EK EK IK
Vm [Ca]
Potassium
Potassium Current
I 50 ms
G.Indiveri (NCS @ INI) ICANN09 Tutorial 50 / 78 G.Indiveri (NCS @ INI) ICANN09 Tutorial 51 / 78
Vmem
Vout
Output voltage
Vmem
Vout
Vmem
time Vout
Reset
Vpw Slope = A
Vpw
Vmem
ICANN09 Tutorial
53 / 78
ICANN09 Tutorial
54 / 78
Capacitive Divider
Positive Feedback
Cfb
voltage
Vout
C2
Vmem
Vout
Vmem
V1
A
C1 V2
Cm
time
V1 =
C2 C1 + C2
V2
Vpw
ICANN09 Tutorial
55 / 78
ICANN09 Tutorial
56 / 78
Gain
How to make voltage gain
Vmem
Vout
Cm tH Ir Vpw tL
time
A
Vdd
tL =
Cfb + Cm Iin
Vmem =
Cfb Iin
Vdd
tH =
Cfb + Cm Ir Iin
Vmem =
Frequency Iin
G.Indiveri (NCS @ INI)
Power Dissipation
The Axon-Hillock circuit is very compact and allows for implementations of dense arrays of silicon neurons
Conductance-based models
Integrate and Fire vs Hodgkin-Huxley
BUT
it has a major drawback: power consumption During the time when an inverter switches, a large amount of current ows from Vdd to Gnd.
Conductance-based (R-C)
ICANN09 Tutorial
59 / 78
ICANN09 Tutorial
61 / 78
Conductance-based models
Integrate and Fire vs Hodgkin-Huxley
But recently proposed models bridge the gap between the two:
Vrest
M6
M1
DPI
M7
M8
Vthr
M2
M3
Leak
Vtau M9 M4
Cahp M10
DPI
Vtau_ahp M13
M17
M21
RefractoryPeriod
ICANN09 Tutorial
62 / 78
ICANN09 Tutorial
63 / 78
Generalized Integrate and Fire models can account for a very large set of DPI neuron sub-threshold equations behaviors captured by far more complicated Hodgkin-Huxley models.
Adaptation
M5 Iin Vahp
SPICE simulations
d dt
M6
umem =
M11 M12
PositiveFeedback in
M14
Cmem
+ F (umem )
20 Vthr=0.325 V Vthr=0.350 V Vthr=0.375 V
14 12 10
M18
15
Leak
Vtau M9 M4
Cahp M10
DPI
Vtau_ahp M13
Imem (A)
M21
Imem (A)
M17
10
8 6 4 2
RefractoryPeriod
0 2
d dt
Imem + Imem 1 I
1
Ig Iin I
+ (Imem )
6 Time (ms)
10
20
40 60 Time (ms)
80
100
+ I0 1 ,
2 + 1
+1
ICANN09 Tutorial
64 / 78
ICANN09 Tutorial
66 / 78
Experimental results
Single spike
Experimental results
Population activity
30 25 Neurons 20 15 10 5 30 25 20 15 10 5 0 0.5 1 1.5 2 0 0 0.5 1 1.5 2
30
25 Neurons 20 15 10 5
0.02
0.04
0.05
67 / 78
ICANN09 Tutorial
Vmem (V)
0.6
0.4
Encode
0.2 0 0 0.05 0.1 Time (s) 0.15 0.2
Decode
1 2 3
1 2 3
Networks of I&F neurons with adaptation, refractory period, etc. Synpases with realistic temporal dynamics Winner-Take-All architectures Spike-based plasticity mechanisms
ICANN09 Tutorial
72 / 78
ICANN09 Tutorial
73 / 78
239
180
A minimum-size chip implementing a recongurable AER neural network. Neurons and synapses have realistic temporal dynamics. Local circuits at The basic each synapse implement theproblem with these models is, of course, generalization: bi-stable a look-up table cannot deal with new events, such as viewing a face from the side rather than the front, and it cannot learn in the predicspike-based plasticity mechanism. One of the simplest and most powerful tive sense described earlier.
Indiveri, Fusi, 2007
types of algorithm developed within learning theory corresponds to networks that combine the activities of units, each broadly tuned to one of the examples (Box 1). Theory (see references in Box 1) shows that a combination of broadly tuned neurons those that respond to a variety of stimuli, although at sub-maximal firing rates might generalize well by interpolating among the examples. In visual cortex, neurons with a bell-shaped tuning are common. Circuits in infratemporal cortex and prefrontal cortex, which combine activities of neurons in infratemporal cortex tuned to different objects (and object parts) with weights learned from experience, may underlie several recognition tasks, including identification and categorization. Computer models have shown the plausibility of this scheme for visual recognition and its quantitative consistency with many data from physiology and psychophysics25 . Figure 2 sketches one such quantitative model, and summarizes a set of basic facts about cortical mechanisms of recognition established over the last decade by several physiological studies of cortex68. Object recognition in cortex is thought to be mediated by the ventral visual pathway running from primary visual cortex, V1, over extrastriate visual areas V2 and V4 to the inferotemporal cortex. Starting from simple cells in V1, with small receptive fields that respond preferably to oriented bars, neurons along the ventral stream show an increase in receptive field size as well as in74 /complexity of their preferred stimuli. the 78 At the top of the ventral stream, in the anterior inferotemporal cortex, neurons respond optimally to complex stimuli such as faces and other objects. The tuning of the neurons in anterior inferotemporal cortex probably depends on visual experience919. In addition, some neurons show specificity for a certain object view or lighting condition13,18,2022. For example, Logothetis et al.13 trained monkeys to perform an object recognition task with isolated views of novel three-dimensional objects (paperclips; Fig. 1). When recording from the animals' inferotemporal cortex, they found that the great majority of neurons selectively tuned to the training objects were view-tuned (see Fig. 1) to one of the training objects. About one tenth of the tuned neurons were view-invariant, consistent with an earlier computational hypothesis23.
NATURE | VOL 431 | 14 OCTOBER 2004 | www.nature.com/nature
Figure 1 Tuned units in inferotemporal cortex. A monkey was trained to recognize a three-dimensional paperclip from all viewpoints (pictured at top). The graph shows tuning to the multiple parameters characterizing each view summarized in terms of spike rate versus rotation angle of three neurons in anterior inferotemporal cortex that are view-tuned for the specific paperclip. (The unit corresponding to the green tuning curve has two peaks to a view of the object and its mirror view.) A combination of such view-tuned neurons (Fig. 2) can provide view-invariant, object specific tuning as found in a small fraction of the recorded neurons. Adapted from Logothetis et al.13.
which several models have been suggested24; see also review in this issue by Abbott and Regehr, page 796), since it depends only on passive experience of the visual inputs. However, the weights of the combination (see Fig. 3) depend on learning the task and require at least some feedback (see Box 2). Thus, generalization in the brain can emerge from the linear combination of neurons tuned to an optimal stimulus effectively defined by multiple dimensions25,23,26. This is a powerful extension of Analog processing, asynchronous digital the older computation-through-memory models of vision and motor control. The question now is whether the available evidence supports the existence of a similar architecture underlying generalization in domains other than vision.
AER INPUT Y
PFC
Categ.
Ident.
AER OUTPUT
AIT
IT
V4/PIT
Technology: Size: Neurons: AER plastic synapses: AER non-plastic synapses Dendritic tree multiplexer:
V4
V1
V1
ICANN09 Tutorial
Summary
Vg Vs Ir Vs Ir
Qs
Figure 2 A model of visual learning. The model summarizes in quantitative terms other models and many data about visual recognition in the ventral stream pathway in cortex. The correspondence between the layers in the model and visual areas is an oversimplification. Circles represent neurons and arrows represent connections between them;G.Indiveri (NCS @ INI) same type. Stages of neurons the dots signify other neurons of the with bell-shaped tuning (with black arrow inputs), that provide example-based learning and generalization, are interleaved with stages that perform a max-like operation3 (denoted by red dashed arrows), which provides invariance to position and scale. An experimental example of the tuning postulated for the cells in the layer labelled inferotemporal in the model is shown in Fig. 1. The model accounts well for the quantitative data measured in view-tuned inferotemporal cortex cells10 (J. Pauls, personal communication) and for other experiments55. Superposition of gaussian-like units provides generalization to three-dimensional rotations and together with the soft-max stages some invariance to scale and position. IT, infratemporal cortex, AIT, anterior IT; PIT, posterior IT; PFC, prefrontal cortex. Adapted from M. Riesenhuber, personal communication.
ICANN09 Tutorial
AER INPUT X
75 / 78
Vd If Vg Vd If
Qd
http://ncs.ethz.ch/
769
Adaptation
M5 Iin Vahp M11 M12
PositiveFeedback
M14
M6
Vrest
M1
DPI
M7
M8
Vthr
M2
M3
Leak
Vtau M9 M4
Cahp M10
DPI
Vtau_ahp M13
M17
M21
RefractoryPeriod
ICANN09 Tutorial
77 / 78
ICANN09 Tutorial
78 / 78