Professional Documents
Culture Documents
Eric Vittoz CSEM Centre Suisse d'Electronique et de Microtechnique SA CH-2007 NEUCHATEL, Switzerland
n n n n
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 information n Strategies n Architectures n 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.
Switch (only function used in digital). Voltage-controlled conductance. Voltage-controlled current source. Multiplier (I D ~ VD x VG + terms). Functions x 2, sqrt(x) (strong inversion: IDsat ~ VG2 + terms). Functions e x, ln(x) (weak inversion: I Dsat ~ eVG/nUT ).
Current memory (open or floating gate). Bipolar operation: - precise e x, ln(x) - voltage and temperature references. Light sensor: any junction. Can only be exploited in analog circuits.
Point of view of power consumption: analog interfaces digital is more efficient for traditional processing
Digital:
An
al
og
Collective processing in massively parallel systems. n Goal: - make a decision - control an action
n
30
SNR [dB]
TRANSLINEAR CIRCUITS
[Gilbert, 1975] n
VBEi
ccw
VBEi
with: Ii =
VBEi Isi e UT
Ii ccwISi ccw
Ii cw ISi cw
=
With MOS transistors in weak inversion: i poor precision acceptable for perceptive processing.
Iout =
Ii
+
I1
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
PSEUDO-CONDUCTANCE/RESISTANCE Linear behaviour of transistors with respect to currents [Bult, 1992] n General case (any current level): same gate voltage I I1 I I2 I1 I2 G1 G2 G2 * for currents G1*
n
conductances G i
n
pseudo-conductances G i*~W/L
Special case of weak inversion ("diffusor" [Boahen, 1992]): i different values of gate voltages possible, thus... i control of G i* by gate voltage, or by... i control current IC: G*~ IC IC Thus: any network of linear controlled resistors can be implemented by transistors only.
ARCHITECTURES
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...
Collective computation of all cells on the best solution - dense interconnections(10 2 to 105 synapses/neuron)
Normalization
[Gilbert, 1984]
Iini Itot
n
Iouti
i current gain same for all cells i adjusts for: I outi = Itot
cell i
[Lazzaro et al., 1988]
Winner-take-all
i points on largest current: I0 for largest I ini Iouti = 0 for others Iinimax i value of largest current
COMMUNICATION
n n n
High connectivity is needed for collective computation but is a... Major obstacle for implementations in 2-dimensional VLSI. 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.
Common m-wire bus. n Activity of each cell transmitted as... n Frequency of address-coding events:
n
code of address
Features: - priority of access based on activity - no clock: phase of events is kept. Parallel access of all cells to the bus - with arbitration or... - without arbitration (random collisions)[Mortara, 1992]
STRATEGIES
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. Adaptation - in time (high-pass) , to eliminate constant information: novelty detector elimination 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
I0 R(k) k
I0
I0
Schematic for 1-D array Real implementation: i 2-D array i pseudo-conductances R*,G*
G(k) proportional to (k) (local photocurrent) i 1/R(k) proportional to F. (k) If R(k) = 0 then I out(k) =I 0 : global normalisation If R(k) >0, each I 0 distributed across area A ~ L 2 = 1/(RG) ~ F i thus: local normalisation in adjustable area A.
n n
Experimental 35x35 hexagonal array of pixels. Output communication by address-coding events. Checker-board illuminated by light gradient:
0
intensity dB
44dB -20
16dB
input current
-40
10
20 column
30
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
PLACE CODING [Landolt, 1998] Representation of a variable x: set of nodes with preferred values x i i xi particular value x = i (activation grade of node i) i 1 center of gravity 0 x0 x1 x2 x3 x4 x5 x6 x7 x8 x
Computation of z = f(x, y) by network of links: each link is a fuzzy AND gate x Advantages: continuous amplitudes high tolerance to perturbations low power in current mode systematic implementation of any function.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
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 2m process Reacts in 50s with 1mW.
CALCULATION OF 1-DIMENSIONAL MOMENTS I0 Ia x=0 resistive I1 R1 R2 Ij IN Rj Rj+1 RN-1 RN Ib x=1 Mxn = Ib 0* Ia VR pseudo-resistive
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
n n
I0 I2 R1* R2*
Ij Rj*
IN RN*
0* Ib
Ambiance-control in vehicles - presence of sun - sun intensity - sun azimuth and elevation Array of 1365 pixels - organized in polar coordinates - light sensor - weighted copies of photocurrents Two linear pseudo-resistive neworks - radial - angular Reacts in 1ms for 100A Precision better than 15.
n n
n n n
Inspired from rabbits eye Ball with random pattern of dots Illuminated by periodic flashes
Calculation of the 2 comp. of translation of a random pattern of dots. n detection of horizontal (and vertical) edges E : x(or y) p x (or y) E E E E E E time t row of sensors E E E E E time t+t E
n n
array of 75 cells, pitch p =300m n resolution above 800 dpi n graceful degradation for defective cells.
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
output
n
1 3 5 7 9
2 4 6 8 0
........
All cells in same sub-layer have n same set of weights with shifted receptive field thus: shift invariant. All cells in same column have same receptive field. Weights learned during design
n
n n n n n n
100K capacitors ( 50K weights) 6x7mm 2 in 0.5 m process Power consumption: <4mW Speed: 1000 images/s. Equiv. 10 8 multiply and add/sec. Wide range of - fonts - quality of printing Graceful degradation with defects.
ONGOING RESEARCH
FEATURE ENHANCEMENT
[Venier et al., 1997] n n n
Edge enhancement by diffusion in linear network (in retinal layer). Address-coding event communication to... Orientation enhancement by nonlinear unisotropic diffusion: i one cell:
photoreceptors edges enhanced
V+
In
V0o
Iin I0
V60o V120o
Out
vertical 60o
orientation enhanced
CSEM, E. Vittoz/ GE99, Alghero, June 11, 1999 - 1
OCCULO-MOTOR SYSTEM
pris mg rat ing lens retin s a
vi per sual cep tion
sal ien ma cy p
ret
sensor tors mo rs enso s
ina
fixe
dw
ide
-an
gle
len
cur
ta ent s r s
te
te upda
Retina with narrow-field lens delivers high-resolution visual info. i moving optical front end (rotating prisms) to change gaze angle. with wide-angle watches the whole visual field. n Retina i saliency map (degree of interest of various parts in scene).
n n
Chip n Chip
controls saccadic exploration of interesting parts (targets). centers and tracks selected target in closed loop.
R target
rotating prisms n Computes 1,2 =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
160000 transistors, 15000 capacitors on 45mm 2 (0.5m process). n Total power consumption at 5V: 5mW.
r r
target
2 2
I0
In
bump (activity of rule n)
bump
current splitters
A1
Vi Vr
An 1 output Iout
R* R*
R* 2R*
R* R*
n n n n n n n n
Rule array: Supply voltage: Power cons.: Input full scale: Settling time: Accuracy:
Further exploration of existing techniques i identification of new applications Exploration of new aspects: i pulse processing (events in time) i stochastic processing i oscillatory behaviour of arrays of nonlinear cells i ......
EXAMPLE OF PULSE (SPIKE) PROCESSING n Extraction of voice pitch (frequency of amplitude modulation)
i Experimental results:
160 140 spikes/second 120 100 130 100 50
ii iii i ii
i
free-running (f0)
i i ii i i i i i i iiii i i i ii i iii i i
1/f
refractory TR
i iii i i
ii
ii i ii
i i
ii
synchronization f = fM
ii i i
short TR
0 1
11
41 21 31 array position
51
low fM
high fM
Biology-inspired systems based on analog VLSI: n still in infancy n slowly growing Potential already demonstrated by working chips: most of them still exploratory n some have already resulted in innovative products.
n
May later compete with digital computation even when power is not limited n for complex perception tasks.
n