Can Biology Inspire Better Circuit Design? The RF Cochlea as a Case Study
Soumyajit [email protected]
Overview
IntroductionBiologically-inspired systemsThe RF cochleaConclusion
Motivations
Emulation: Biology solves problems that computers have difficulty with
AdaptationPattern recognitionLow-power, real time computation
Computation: Biological models can be simulated faster in hardware
Challenges
Modeling challengesParameter values hard to obtainFidelity hard to verifyFiguring out reasonable simplifications is hard
As computational media, biology and silicon are very differentNeuronal networks are 3D, silicon is planarNeural networks are hybrid state machines
The human auditory periphery
Biological cochlea numbers
Dynamic range 120 dB at inputPower dissipation ~14μW (estimated)
Power supply voltage ~150 mVVolume ~35mm x 1cm x 1 cmDetection threshold at 3 kHz 0.05 Å at eardrumFrequency range 20 Hz – 20 kHzOutlet taps ~35,000Filter bandwidths ~1/3 OctavePhase locking threshold ~5 kHz
Information is reported with enough fidelity so that the auditory system has thresholds forITD discrimination at ~10 μsFreq. discrimination at 2 Hz (at 1kHz)Loudness discrimination ~1 dB
The bottom line
Biology has evolved a broadband spectrum analyzer withExtremely low power consumptionHigh dynamic rangeHigh resolution (~1Hz around 2KHz)
Binaural hearing allowsPrecise arrival time discrimination (to within 10μs)Spatial localization of sound sources
Conventional spectrum analyzers
Essentially a swept-tuned superheterodyne receiverIF filter sets resolution bandwidth (RBW)Sweep time proportional to 1/(RBW)2
Trade-off between speed and precisionSubstantial speedup by using an FFT (instead of an analog IF filter) for small resolution bandwidths
Spectrum analyzers: prior engineering versus biology
Trade-off between speed, precision (number of bins N) and hardware complexity
Topology Acquisition time Hardware complexity
Real time?
FFT O(N log(N)) O(N log(N)) No
Swept-sine O(N2) O(1) No
Analog filter bank O(N) O(N2) Yes
Cochlea O(N) O(N) Yes
The cochlea is an ultra-wideband spectrum analyzer with extremely fast scan time, low hardware complexity and power consumption, and moderate frequency resolution
Example 1: a silicon cochlea
An analog electronic cochlea, Lyon, R.F.; Mead, C.;Acoustics, Speech, and Signal Processing, IEEE Transactions on, Volume 36, Issue 7, July 1988 Page(s):1119 - 1134
The mammalian retina
Example 2: a silicon retina
Silicon retina with correlation-based, velocity-tuned pixels, Delbruck, T.; Neural Networks, IEEE Transactions on, Volume 4, Issue 3, May 1993 Page(s):529 - 541
Example 3: a silicon muscle fiber
An analog VLSI model of muscular contraction, Hudson, T.A.; Bragg, J.A.; Hasler, P.; DeWeerth, S.P.; Circuits and Systems II: Analog and Digital Signal Processing, IEEE Transactions on , Volume 50, Issue 7, July 2003 Page(s):329 - 342
Example 3: a silicon muscle fiber
The human auditory periphery
Structure of the cochlea
The cochlea is a long fluid-filled tube separated into three parts by two membranesHuman cochleas are about 3.5mm long
Coiled into 3.5 turns to save space1mm in diameter
Oval and round windows couple sound in and outFluid – membrane interactions set up traveling wave from base to apex
Cross-section of the cochlea
Cochlea powered by ionic gradient between perilymph and endolymph
Provides a quiet power supply isolated from blood circulation
Basilar membrane Supports traveling waveSupports organ of Corti
Reissner’s membrane has no mechanical functionInterface with 25,000 endings of the auditory (eighth cranial) nerve
Perilymph
Perilymph
Endolymph
Organ of Corti
Contains mechanisms forSignal transduction (inner hair cells)Active cochlear amplification (outer hair cells)Neural coding of auditory information (spiral ganglion cells)
Stereocilia (hairs) used for sensingActuation and amplification mechanism unclear
The basilar membrane
Properties of basilar membrane change (taper) exponentially with position (from base to apex)
Width increases (from 50 to 500μm)Stiffness decreases
Hence resonant frequency of the fluid – membrane system also depends exponentially on position along the cochlea
Spectral analysis!
Wave motion
Tonotopic map: exponential scaling
Frequency–to–place transform
Cochlear frequency responses
Frequency responses of live cochleas are sharper & have more gain Implies presence of an active cochlear amplifierSpatial responses look very similar to frequency responses (frequency-to-place transform)
Gain control
Strong compressive nonlinearity present in cochlear response with sound levelEffects of compressive gain control
Enhanced dynamic rangeTwo-tone suppression (masking)
Models of cochlear damping versus local signal amplitude |A|
Experimental cochlear frequency responses versus input amplitude (sound pressure level (SPL) in dB)
( ) 1 1 logd A Aλ σ≡ + ⋅
( ) 2 2d A Aλ σ≡ + ⋅
( ) 23 3d A Aλ σ≡ + ⋅
“log law”
“power of 1 law”
“power of 2 law”
Gain control (continued)
Simple model: feedback loop with compressive nonlinearityBehavior
Linear at small and large amplitudesStrongly compressive in between
Beyond the cochlea
10 nerve endings per inner hair cell~20dB dynamic range in firing rate per nerve fiberSmart neural coding to increase total output dynamic range
The auditory pathway
Auditory nerve connections in the cochlea
Why an RF cochlea?
Silicon cochleas have been built at audio frequencies, but operating at RF has several advantages
Availability of true (passive) inductors at RF frequenciesReduced noise
Improved performance because of new theoretical insightsSeveral possible applications
Fast, wideband real-time spectrum analysisFront end for wideband radio receiversAs a distributed “RF laser”
Proposed implementationOperating frequency range
8GHz – 800MHz (bidirectional)6GHz – 450MHz (unidirectional)
Over 60dB of input dynamic range
Cochlear models
Fluid mass modeled as network of inductors or resistorsBasilar membrane modeled by complex impedanceSimplifications
1D models: if a single propagating wave mode is considered A cascade of unidirectional filters: if reflected waves are ignored
One dimensional models
Two dimensional model
Bidirectional RF cochlea
RF cochlea chip die photos
Unidirectional
Bidirectional
Spatial responses
5 10 15 20 25 30 35 40 45-70
-60
-50
-40
-30
-20
-10
0
Stage Number
Out
put v
olta
ge (d
B)
8 GHz
1GHz
Two-tone responses
Sta
ge n
umbe
r
20 40 60 80 100
5
10
15
20
25
30
35
40
45 -70
-60
-50
-40
-30
-20
-10
Varying the negative resistance
0 10 20 30 40-70
-60
-50
-40
-30
-20
-10
0
Stage number
Out
put v
olta
ge (d
B) 1.5 GHz
2.3 GHz3.5 GHz
8 GHz5.3 GHz
Driving the cochlea unstable
Active element bias (V)
Freq
uenc
y (G
Hz)
0.6 0.65 0.7 0.75 0.8
1
2
3
4
5
6
7
80
5
10
15
20
25
A video of the RF cochlea in action
Faculty members in related areas
Harvard-MIT division of Health Sciences and Technology (HST)Prof. Dennis Freeman (Cochlear micro-mechanics)Profs. Christopher Shera, Bertrand Delgutte and Donald Eddington (Auditory physics)Prof. Roger Mark (Modeling & control of complex physiological systems)
Profs. Joel Voldman & Jongyoon Han (BioMEMS)Prof. Rahul Sarpeshkar (Analog VLSI and biological systems)Prof. Joel Dawson (Biomedical circuits and systems)Prof. George Verghese (Modeling and control of complex physiological systems)Prof. Scott Manalis (Nanoscale sensing)Many others ...
Other info
Useful classesCircuit design: 6.101, 6.301, 6.331, 6.374, 6.376, 6.775, 6.776Control systems: 6.011, 6.302, 6.241Bioelectronics: 6.021J, 6.022J, 6.023J, 6.024J, 6.121MEMS: 6.777Biomedical systems: 6.971
Companies of interestImplanted devices: Medtronic, Advanced BionicsBiomedical systems: GE, PhilipsMany others!
Computational Intelligence for Understanding Earth SystemsSai Ravela, MIT EAPS
Tuesday, Dec. 45:30-6:30 PM
Room 34-401A(dinner to follow)
Backup slides
Cochlear models
The definition of the cochlea Transfer Function (TF) is
Bidirectional Cochlear Model
( )dP j L x Udx
ω= − ⋅
( ),dU Pdx Z j xω
= −
( ) ( )( ) ( ) ( ) ( )
1 1,0 0 0 ,
outI x dU PTF j xU U dx U Z j x
ωω
≡ = − =
P – pressure (voltage)U – volume velocity (current)L(x) – liquid mass (inductance)Z(jω, x) – Basilar Membrane (BM)
impedance
The Center Frequency (CF):
where is the CF on the basal end of the cochlea
In the real cochlea the BM impedance Z(jω,x) as well as U, P and TF depend only on the following combination of x and ω:
where is the inductance per unit length on the basal end of the cochlea
is the cochlea taper coefficient
The liquid mass, or inductance, L(x) increases exponentially with position x:
Scaling of the Cochlea
l
(0)cω
0L
/( ) (0)c c
x lx eω ω −= ⋅
( ) 0
/x lL Lx e= ⋅
( ) ( ) ( ) /,0n x l
c c
xx e
ω ωω ωω ω −≡ =
⋅ n ns jω≡
WKB Analytical Solution
( ) ( ) ( )3/ 2
0
expns
n n n n nTF s s k s k s ds⎛ ⎞
′ ′∝ ⋅ ⋅ −⎜ ⎟⎜ ⎟⎝ ⎠∫
• The WKB-approximate solution for the cochlea TF is
( ) ( ) ( )logn n nn n
d dk j Phase TF j TFd d
ω ω ωω ω
≈ − + ⋅
• Ignoring the pre-exponent dependencies,
• Now, by knowing the experimental cochlea collective response, we can calculate k(jωn) and snZn(sn), and therefore design the cochlea section
( )2
22 nn
d P k s Pds
= ⋅ ( ) ( )( ) ( )
2 202 0c
nn n n n n
l L Nk ss Z s s Z sω⋅ ⋅
= ≡⋅ ⋅
• The ODE for the pressure, or voltage, P is
Designing Zn(sn) to be a Rational Function
The simplest possible rational function is
( ) ( )22
2 2
2 1
0.10.763.8
n nn n n
n n
s dss Z s
s sQ
d
Q
μ μ
μ
+ +⋅ =
+ +
===
We tweak these parameters to obtain a desirable cochlea frequency response
Pole-zero diagram of snZn(sn)
Want Z n to be a rational function so that it can be easily implemented
Frequency Response of snZn
Double zero in snZn close to the jωaxis vital for collective gain
snZn close to zero for a range of frequencies around ωn = 1Several stages contribute gain
Real part of Zn < 0 for ωn < 1 Traveling wave amplitude increases before CFZn cannot be completely passive
Modified Cochlear Architectures
Possible modifications(a) Reverse the mechanical – to – electrical mapping convention(b) Use a low pass to high pass (s → 1/s) transformation
Problems(a) Need to synthesize complex floating, bidirectional impedance(b) High frequencies have to travel the whole length of the cochlea
Synthesizing the Cochlear Impedance
Use coupled resonator topology to synthesize Zn
Suitable for IC implementationComputer-based optimization using Mathematica™ used to find component valuesSingle active element required – R1 must be negativeAdditional synthesis constraints
|k| < 0.8 so that an integrated transformer can be usedC1 & C2 > Cmin to absorb parasitic capacitances from inductors and resistors
1 2
MkL L
=
Negative Resistance CircuitsCross-coupled differential pair Inductive gate degeneration
Coupled inductorsCapacitive source degeneration
Problem: these circuits cannot synthesize floating negative resistors
Cochlear Transfer Functions
Input impedance of the cochleaResistive over the operating frequency rangeReactive otherwise
Frequency scaling
Impedance scaling
Spatial transfer functions
Termination Issues
Instabilities due to reflections from Apical terminationInter-stage impedance mismatch
Causes spontaneous oto-acoustic emissions (SPOAE’s) in biological cochleasSimilar to how a laser worksReduce apical reflections by using a perfectly matched terminating layer (PML)
System eigenvalues with (A) single terminating impedance (B) distributed terminal layer
Unidirectional Cochlea with Improved Section TF
( )
( ) ( )( )
( ) ( )
1
1
1
exp
exp
11
n
n
s
ns
n n n n
nn n n
TF k s ds
TF k s s s
TFk s s s
−
−
−
⎛ ⎞= −⎜ ⎟⎜ ⎟
⎝ ⎠≈ − ⋅ −
≈+ ⋅ −
∫
• The TF of the n-th section is
• The TF of the n-th section of the cochlea with Noct sections per octave is
( )( )
,
,2
,
1ln 21
2 1
out n
n nin n
oct n out n n
Vs sV N
N s d V sμ
=⋅ +
+ ⋅+ ⋅ +
Unidirectional Cochlea with Improved Section TF
( ) ( ) ( )3/ 2
0
expns
n n n n nTF s s k s k s ds⎛ ⎞
′ ′∝ ⋅ ⋅ −⎜ ⎟⎜ ⎟⎝ ⎠∫
( ) ( )11
expj
j
sn
nj s
TF s k s ds−
=
⎛ ⎞⎜ ⎟= −⎜ ⎟⎝ ⎠
∏ ∫
( )1
expn
n
s
ns
TF k s ds−
⎛ ⎞= −⎜ ⎟⎜ ⎟
⎝ ⎠∫
• Ignoring the pre-exponent dependencies,
• Already looks like a cascade of filters, with the TF of the n-th section being
• The WKB-approximate solution for the cochlea TF is
Action of a filter cascade
Preliminary specifications for the RF cochlea
Parameter Unidirectional Bidirectional
Fabrication technology UMC 0.13µm CMOS UMC 0.13µm CMOS
Maximum input signal 700mVrms 700mVrms
12 (17 / e-fold)
50
7GHz – 400MHz
~ 5
~ 20dB
< 2mVrms
71dB
Input impedance 50Ω 50Ω
Maximum scan clock speed 10MHz 10MHz
75mA @ 1.0V
Stages per octave 14 (20 / e-fold)
Number of stages 50
Frequency range 9GHz – 800MHz
Transfer function Q3dB 15
Transfer function gain 0dB
Output noise < 300µVrms
Input-referred dynamic range 67dB
Power consumption 120mA @ 1.5V
‘Traditional’ software radio consumes 7W just for a 9-bit, 10GHz ADC
Frequency responses
100
-70
-60
-50
-40
-30
-20
-10
0
Frequency (GHz)
Out
put v
olta
ge (d
B)
Stage 46
Stage 6
Compression curves
-60 -50 -40 -30 -20 -10 0-70
-60
-50
-40
-30
-20
-10
0
Input power level (dBm)
Out
put v
olta
ge (d
B)
fmax
fmax/1.5
fmax/2.3
fmax/3.5
fmax/5.3
Varying the line loss cancellation
0 10 20 30 40-70
-60
-50
-40
-30
-20
-10
0
Stage number
Out
put v
olta
ge (d
B)
3.0 GHz
1.3 GHz