neuromorphic enginerring- subthrshold design
DESCRIPTION
Neuromorphic Enginerring- subthrshold designTRANSCRIPT
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
BIO-INSPIRED ANALOG VLSI CIRCUITS
AND THEIR INDUSTRIAL APPLICATION
Eric VittozCSEM Centre Suisse d'Electronique et de Microtechnique SA
CH-2007 NEUCHATEL, Switzerland
n Biology-inspiration
n Examples of industrial application
n Ongoing research
n Conclusion.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
BIOLOGY AS A SOURCE OF INSPIRATION
In the continous fight for survival all along evolution:
n life has invented increasingly complex solutions
Biology inspired systems:
n borrow some of them
n adapt to constraints of available technologies
Inspiration for VLSI processing circuits: from the brain
Different levels:
n Representation of signals and informationn Strategiesn Architecturesn Methodology
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
METHODOLOGY
n Opportunism:
Along evolution, life has made, step after step:- best possible use of existing strucures as...- bootstrap for the next level of compexity.
Bottom-up approach, can be applied to VLSI:
- identify all properties of technologies, devices, circuits
- exploit them opportunistically to build more complex systems.
Also:
- explore potential of a solution in a broader range
- search for hidden advantages of apparent drawbacks.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
POSSIBLE FUNCTIONS OF MOS TRANSISTORS
OPPORTUNISM:
n Switch (only function used in digital).
n Voltage-controlled conductance.
n Voltage-controlled current source.
n Multiplier (I D ~ VD x VG + terms).
n Functions x 2, sqrt(x) (strong inversion: IDsat ~ VG2 + terms).
n Functions e x, ln(x) (weak inversion: IDsat ~ eVG/nUT ).
n Current memory (open or floating gate).
n Bipolar operation:- precise e x, ln(x)- voltage and temperature references.
n Light sensor: any junction.
Can only be exploited in analog circuits.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
PROSPECTS FOR ANALOG IN SIGNAL PROCESSING
0 30 60 90 120
10-9
10-12
10-15
10-18
SNR [dB]
Analog
Pmin/f [J]
Digital:
digital is moreefficient for “traditional”
processing
analog interfaces
(also, qualitatively: precision, linearity)
(also, qualitatively: chip area)
analogfor
perception n Goal: - make a decision- control an action
n Collective processing in massively parallel systems.
n Point of view of power consumption:
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
Π ISi
Π ISi
Π Ii
Π Ii
cw cw
ccwccw
= = λ
Ii
VBEi
n With bipolar transistors:
ΣVBEi = ΣVBEicw ccw
Ii = Isi eVBEiUTwith:
thus: VBEi = UT lnIiIsi
which yields:
cw = clockwiseccw = counter-clockwise
TRANSLINEAR CIRCUITS
n With MOS transistors in weak inversion:
i poor precision acceptable for perceptive processing.
[Gilbert, 1975]
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
EXAMPLE OF TRANSLINEAR CIRCUITS
VECTOR-LENGTH CALCULATION
I1
-
[Gilbert, 1975]
Iout = Ii2Σ
I1
+ + +
+p-MOS in weak inversion
Current mode, provide:
npn bipolar
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
PSEUDO-CONDUCTANCE/RESISTANCE
G1 G2
I1 I2I
G1* G2*I1 I2
I
for currents
n Linear behaviour of transistors with respect to currents [Bult, 1992]
n General case (any current level): same gate voltage
n Special case of weak inversion ("diffusor" [Boahen, 1992]):
i different values of gate voltages possible, thus...
i control of Gi* by gate voltage, or by...
i control current IC:
n Thus: any network of linear controlled resistors can be implemented by transistors only.
ICG*~ IC
conductances Gi pseudo-conductances Gi*~W/L
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
n Massive parallelism (10 11 neurons in brain)
VLSI technologies: towards billion-transistor chips
- large arrays (102-106) of...
- simple cells (10 - 103 tr.)
- low local speed (< 10kHz)
- limited local precision thanks to...
n Collective computation
- of all cells on the best solution
- dense interconnections(102 to 105 synapses/neuron)
ARCHITECTURES
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
EXAMPLES OF COLLECTIVE OPERATORS
n Normalization [Gilbert, 1984]
Σ Iouti = Itot
cell iItot
IoutiIini
n Winner-take-all [Lazzaro et al., 1988]
I0+
-
Iini
cell i
IoutiIinimax
i current gain same for all cells
i adjusts for:
Iouti =I0 for largest Iini0 for others
i points on largest current:
i value of largest current
n Consensus achieved by communication through single wire.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
COMMUNICATION
n High connectivity is needed for collective computation but is a...
n Major obstacle for implementations in 2-dimensional VLSI.
n Possible (complementary) solutions:
i collective computation via single or few wire(s)
i communication with nearest neighbours only
i hierarchical interconnections (smaller density for larger distance)
i analog multiplexing:
- standard row/column scanning
- asynchronous and event driven
- pulse communication.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
COMMUNICATION BY ADDRESS-CODING EVENTS
n Common m-wire bus.n Activity of each cell transmitted as...n Frequency of address-coding events:
code ofaddress
m-wirebus
event (set of very short pulses)
n Parallel access of all cells to the bus- with arbitration or...- without arbitration (random collisions)[Mortara, 1992]
n Features: - priority of access based on activity- no clock: phase of events is kept.
[Mahowald, 1992]
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
n Learning from examples (instead of programming)- on-chip learning not very useful in practice but...- learning on computer during design phase
resulting synaptic weights dumped on the chip.
n Adaptation
- in time (”high-pass”) , to eliminate constant information:novelty detectorelimination of constant spatial noise of sensors.
- to signal level, to cover wide dynamic range
140 dB of external excitation
40 dB of nerve activation
very useful for analog VLSI processing
STRATEGIES
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
I0 I0 I0
k
Iout(k)
G(k)
R(k)
(k)
F. (k)
pixel k-1 pixel k pixel k+1
RETINA FOR LOCAL ADAPTATION
n G(k) proportional to (k) (local photocurrent)
[Venier, 1997]
n If R(k) = 0 then ΣI out(k) =ΣI0 : global normalisation
n If R(k) >0, each I 0 distributed across area A ~ L2 = 1/(RG) ~ F
i thus: local normalisation in adjustable area A.
n Schematic for 1-D array
n Real implementation:
i 2-D array
i pseudo-conductances R*,G*
i 1/R(k) proportional to F. (k)
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
0
-20
-40
10 30
inte
nsi
ty d
B 44dB 16dB
20column
input current
output current
n Checker-board illuminated by light gradient:
n Experimental 35x35 hexagonal array of pixels.
n Output communication by address-coding events.
MEASUREMENTS ON LOCALLY ADAPTIVE RETINA
n Dynamic range reduced to that of the object contrast.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
n Pulse representation (in biology: spikes of action potential)
- for communication
- for processing (audition...ubiquitous?)
time-domain processing
degree of coincidence of events
synchronization of processes
n "Place coding" (patterns of activity in 2D maps)
REPRESENTATION OF SIGNALS AND INFORMATION
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
PLACE CODING
Representation of a variable x:set of nodes with preferred values xi
0
1µi (activation grade of node i)
xx0 x1 x2 x3 x4 x5 x6 x7 x8
particular value
x y z
µi xiµi
x = ΣΣ
center of gravity
Computation of z = f(x, y) by network of links: each link is a fuzzy AND gate
Advantages:continuous amplitudes
high tolerance to perturbations
low power in current mode
systematic implementation of
[Landolt, 1998]
any function.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
EXAMPLES OF
INDUSTRIAL APPLICATIONS
Biology-Inspired Analog VLSI:
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
DETECTION OF MAXIMUM SPOT INTENSITY
Control of LCD protection screen
Array of 26x26 pixels- light sensor- element of max.-current copier
18V generator, control and test
4x4.5 mm2 in 2µm process
Reacts in 50µs with 1mW.
[Chevroulet et al., 1995]
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
CALCULATION OF 1-DIMENSIONAL MOMENTS
R1
I1R2 Rj
IjRj+1 RN-1 RN
INI0
Ia Ibresistive
VR
RN*
IN
Ib
R1*
I2R2*
I0
Ia
IjRj* 0*0*
pseudo-resistive
x=0 x=1
n If: R j = jn-1 R1
n Then, n th order momentof current distribution:
Mxn = Ib
n Center of gravity x 0:
i n = 1 (Rj constant)
i x0 = Mx/M0 =Ib
Ia+ Ib
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
MONITORING OF SOLAR ILLUMINATION[Venier et al., 1996]
n Ambiance-control in vehicles- presence of sun- sun intensity- sun azimuth and elevation
n Array of 1365 pixels- organized in polar coordinates- light sensor- weighted copies of photocurrents
n Two linear pseudo-resistive neworks- radial- angular
n Reacts in 1ms for 100µA
n Precision better than 15°.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
MOTION DETECTOR FOR TRACKBALL
n Inspired from rabbit’s eye
n Ball with random pattern of dots
n Illuminated by periodic flashes
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
E E EE E EE E EE E E
time ttime t+∆t
x (or y)p∆x(or ∆y)
row of sensors
MOTION DETECTOR FOR POINTING DEVICE
n Calculation of the 2 comp. of translation of a random pattern of dots.n detection of horizontal (and vertical) edges E :
n identification of edges moving during ∆t (between 2 snapshots)
n Motion during ∆t:
n current mode analog processing
n array of 75 cells, pitch p =300µm
n resolution above 800 dpi
n graceful degradation for defective cells.
∆x (or ∆y) = p*number of moving edges E
total number of edges
[Arreguit et al.,1996]
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
OPTICAL CHARACTER RECOGNITION
67 89 0
1 23 45........
output[Masa et al., 1999]
n Convolutive neural network.
n abstraction level increases at each step.
n All cells in same sub-layer haven same set of weights with shifted receptive field
thus: shift invariant.
n All cells in same column have samereceptive field.
n Weights learned during design
n implemented as capacitance ratios
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
THREE-LAYER CONVOLUTIONAL NEURAL NETWORK
n 100K capacitors ( 50K weights)
n 6x7mm 2 in 0.5 µm process
n Power consumption: <4mW
n Speed: 1000 images/s.
n Equiv. 10 8 multiply and add/sec.
n Wide range of- fonts- quality of printing
n Graceful degradation with defects.
[Masa et al., 1999]
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
ONGOING RESEARCH
Biology-Inspired Analog VLSI:
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
edgesenhanced
vertical
orientation enhanced
FEATURE ENHANCEMENT
n Edge enhancement by diffusion in linear network (in retinal layer).
n Orientation enhancement by nonlinear unisotropic diffusion:
n Address-coding event communication to...
Out
InIin
V+
V60oV0o
I0
V120o
i preferred orientation electrically adjustable.
60o
photo-receptors
i one cell:
[Venier et al., 1997]
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
current state
updates
retina
fixed wide-angle lens
prism gratings saliencymap
visualperceptionlens
sensor
motors
sensors
retina
saccadescontrolTrackingcontrol
final user orfurther processing
OCCULO-MOTOR SYSTEM
n Retina with narrow-field lens delivers high-resolution visual info.
i moving optical front end (rotating prisms) to change gaze angle.n Retina with wide-angle watches the whole visual field.
i saliency map (degree of interest of various parts in scene).
n Chip controls saccadic exploration of interesting parts (targets).n Chip centers and tracks selected target in closed loop.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
CHIP FOR SACCADES CONTROL[Landolt, 1997]
ϕ1 ϕ2
rotating prisms
R
n
n n
n
target
visual field of retina
n Computes ϕ 1,ϕ2 =f(R) to center visual field of retina on target.
n Exploration time of targets function of their degree of interest n nn n
n Place-coding + communication by address-coding events.
n 160’000 transistors, 15’000 capacitors on 45mm 2 (0.5µm process).n Total power consumption at 5V: 5mW.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
ϕ1 ∆ϕ1ϕ2
∆ϕ2
rotating prisms
CHIP FOR VISUAL TO MOTOR MAPPING
n Incremental control: computes ∆ϕ 1,∆ϕ2 = f(ϕ1,ϕ2, r) needed for ∆r
n Place coding; activation grade represented by :
i current (internal): 0...100nA
i pulse frequency (external): 0...1KHz
n Total power consumption at 5V: 30µW
nr∆r
visual fieldof retina
target[Landolt, 1997]Area: 2mm2
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
2R*
R*Iout
A1 An
bumpVi11Vr11
bumpVin1Vr1n
bumpVinm
Vrnmbump
Vi1m
Vr1m
I0
I1 In
n rules
m in
pu
t va
riab
les
~MIN
current splitters
4-bit weightingof activities
(activity of rule n)
1 outputVi Vr
+
ANALOG FUZZY CONTROLLER[Landolt, 1996]
R* R*
R* R*
bump circuit(input membership function)
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
FUZZY CONTROLLER CHIP[Landolt, 1996]
n Rule array: 8 by 10
n Supply voltage: 1.8V
n Power cons.: 850nW
n Input full scale: 1V
n Settling time: 0.5ms
n Accuracy: 2.6% RMS
n CMOS process: 2µm
n Chip core area: 1.2x1.1mm 2
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
n Further exploration of existing techniques
i identification of new applications
n Exploration of new aspects:
i pulse processing (events in time)
i stochastic processing
i oscillatory behaviour of arrays of nonlinear cells
i ......
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
i i iii i
i
i
iii
i
i i i i iiii i
iii i
ii i i
i
ii i i i ii
iii i i
ii i
i
i i ii i
0
50
100
130
iiiii
i i iii ii i i i
i ii i i i
iiiiii i i
ii i
iii i i i
iii i i
i ii i
i
i
ii
ii i ii
100
120
140
160
1 11 21 31 41 51array position
spik
es/
seco
nd
free-running (f0)
synchronization f = fM
output of coincidence array(fM = 130Hz)
i Experimental results:
EXAMPLE OF PULSE (SPIKE) PROCESSING
n Extraction of voice pitch (frequency of amplitude modulation)
1/f0
1/f
integration
effect of input
pulse
i Behaviour of “chopper cell”:
i System architecture [van Schaik, 1998]:silicon cochlea
1/fM
low fM high fM
long TR short TRchopper array
coincidence detector arrayi Application to voice separation?
refractory TR
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
n Biology-inspired systems based on analog VLSI:n still in infancyn slowly growing
n Potential already demonstrated by working chips:
n most of them still exploratoryn some have already resulted in innovative products.
n Very attractive for low-power smart microsystems
n in particular for low-level vision.
n May later compete with digital computation
n even when power is not limitedn for complex perception tasks.