A Silicon Cochlea With Active Coupling , Member, IEEE

A Silicon Cochlea With Active Coupling , Member, IEEE
444
IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 6, DECEMBER 2009
A Silicon Cochlea With Active Coupling
Bo Wen, Member, IEEE, and Kwabena Boahen, Member, IEEE
Abstract—We present a mixed-signal very-large-scale-integrated chip that emulates nonlinear active cochlear signal processing. Modeling the cochlea’s micromechanics, including outer
hair cell (OHC) electromotility, this silicon (Si) cochlea features
active coupling between neighboring basilar membrane (BM)
segments—a first. Neighboring BM segments, each implemented
as a class AB log-domain second-order section, exchange currents
representing OHC forces. This novel active-coupling architecture
overcomes the major shortcomings of existing cascade and parallel filter-bank architectures, while achieving the highest number
of digital outputs in an Si cochlea to date. An active-coupling
architecture Si cochlea with 360 frequency channels and 2160
pulse-stream outputs occupies 10.9 mm2 in a five-metal 1-poly
0.25- m CMOS process. The chip’s responses resemble that of
a living cochlea’s: Frequency responses become larger and more
sharply tuned when active coupling is turned on. For instance,
gain increases by 18 dB and 10 increases from 0.45 to 1.14. This
enhancement decreases with increasing input intensity, realizing
frequency-selective automatic gain control. Further work is required to improve performance by reducing large variations from
tap to tap.
Index Terms—Class AB, cochlear amplifier, log-domain, neuromorphic, silicon (Si) cochlea.
I. SILICON COCHLEAE
ILICON (Si) cochleae emulate cochlear processing of
sound stimuli in very-large scale integrated (VLSI) systems, attempting to match the biological cochlea’s sound
sensitivity, frequency selectivity, and dynamic range. The effort
to build artificial cochleae in Si has been largely motivated by
their potential applications in hearing aids, cochlear implants,
and other portable devices that demand real-time, low-power
signal processing for speech recognition; these requirements
favor subthreshold analog VLSI designs [1]. Furthermore,
analog VLSI is amenable to cochlea-like distributed processing
due to its compact computational elements, large numbers of
which can be integrated in a small area of Si. However, digital
outputs are easier to interface with higher level processing,
whether performed by other neuromorphic chips or implemented in computer software. Thus, a mixed-mode approach,
where the cochlea’s analog outputs are converted to digital
pulses—a function performed by the auditory nerve (AN)—is
most attractive.
S
Manuscript received August 17, 2008; revised November 21, 2008. First published October 30, 2009; current version published November 25, 2009. This
work was supported by the Packard Foundation, under Grant 99-1454. This
paper was recommended by Associate Editor T. Delbruck.
B. Wen was with the Department of Bioengineering, University of Pennsylvania. She is now with the Research Laboratory of Electronics, Massachusetts
Institute of Technology, Cambridge, MA 02139 USA (e-mail: [email protected]).
K. Boahen was with the University of Pennsylvania, Philadelphia, PA 19104
USA. He is now with the Department of Bioengineering, Stanford University,
Stanford, CA 94305 USA (e-mail: [email protected]).
Digital Object Identifier 10.1109/TBCAS.2009.2027127
Si cochleae take the form of a bank of low-pass or band-pass
filters, with exponentially decreasing resonant frequencies, connected in cascade or in parallel. Cascaded filter banks, introduced in the first Si cochlea [2], rely on gain accumulation, with
each filter’s gain being small. Their major drawbacks are excessive delay and noise accumulation [3], and poor fault tolerance.
Parallel filter banks require each filter to generate the desired
gain and tuning by itself, falling short of the biological cochlea’s
frequency tuning and cutoff slopes [4]. A variation of the parallel architecture introduced by Watts [5] couples the filters together through a resistive grid that models the cochlear fluid.
Although this coupled architecture emulates the cochlea more
faithfully, its gain is diminished by destructive interactions [6].
Our Si cochlea aims to overcome existing architectures’
shortcomings by mimicking the cochlea’s micromechanics,
in particular, the intricate anatomical arrangement of outer
hair cells (OHCs) and other structural cells in the organ of
Corti. Although it is a mystery as to how exactly OHC motile
forces, discovered in mammalian cochlea more than two
decades ago [7], boost the basilar membrane’s (BM) vibration,
cochlear microanatomy provides clues. Based on these clues,
we previously proposed a novel mechanism for the cochlear
amplifier—active bidirectional coupling (ABC) [8]. Here, we
report a mixed-signal VLSI chip that implements ABC, the
first cochlear chip that employs active behavior (i.e., negative
damping [9]–[11]) instead of passive behavior (i.e., undamping
[12], [13]1).
By counteracting the coupled architecture’s destructive
interference, ABC promises frequency tuning comparable to
human performance. The psychophysically measured auditory
filter width, or critical band, is about 1/3 to 1/6 octave [14].
—center frequency divided by
This bandwidth suggests a
width 10 dB below the peak—of between 3 and 6 at the BM. In
fact,
values measured from cat AN fibers increase from 1
to above 6 from 200 to 20 kHz [15]. In comparison, the highest
reported for the cascade and parallel architectures are 0.92
[2] and 0.42 [4], respectively. In the former, the individual filter
must be limited to manage noise accumulation [3]; in the
latter, individual filters had to achieve the desired performance
on their own. By avoiding these constraints, the passively
of 2.34—the
coupled architecture achieved a best-case
highest to date—despite a 25-dB gain-drop due to destructive
interference [6]. Our software simulations suggested that ABC
could counteract this destructive interference [8], thereby
achieving performance comparable to humans. However, this
proved challenging in Si: We discovered that reflections were
caused by abrupt changes in BM properties (due to the transistor mismatch).
Section II presents the challenges that VLSI implementations
of ABC face. Section III presents a mathematical model of
Q
1The term “active” used in [13] refers to the fact that the
is actively controlled; this differs from our use of active, which refers to actively pumping
energy into the traveling wave through negative damping.
1932-4545/$26.00 © 2009 IEEE
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
WEN AND BOAHEN: SILICON COCHLEA WITH ACTIVE COUPLING
445
ABC, first proposed in [8]. Section IV presents the synthesis of
an analog circuit that satisfies the model’s equations. Section V
presents a transistor-level circuit implementation. Section VI
presents real-time chip responses that emulate nonlinear active
cochlear behavior. Section VII discusses the impact of transistor mismatch. Section VIII concludes this paper. This paper
extends the work described in [9] and [16] .
II. IMPLEMENTATION CHALLENGES
While our software simulations demonstrated ABC’s
promise as a cochlear amplifier [8], [17], implementing it in
digital or analog VLSI presents challenges. According to our
simulations, a large gain (more than 60 dB) is achieved when
negative damping (i.e., active amplification) occurs over many
BM segments (about 60). This requirement necessitates a large
number of segments per octave (about 45) if sharp tuning
) is desired as well. The upshot shows that about
(
450 segments are needed to span the audio-frequency range
(20–20 kHz or ten octaves), presenting challenges for digital
and analog implementations.
As for digital VLSI, although the bit-serial technique
offers implementation efficiency, it is hard-pressed to fit several hundred segments on a chip. This approach yielded 71
second-order sections in a 40 mm 1.2- m-complementary
metal–oxide semiconductor (CMOS) application-specific integrated circuit (ASIC) [18] and 88 sections in a Xilinx Virtex
XCV1000 FPGA [19]. Extrapolating these numbers yields 409
sections in a 10 mm 0.25
-CMOS ASIC and 334 sections
in a Xilinx Virtex II XC2V8000, but this does not include the
fluid model nor does it include ABC.2 Adding this functionality,
which requires two multiply-accumulates for ABC and about
ten for the fluid (per section), will double the complexity of the
system, and halve the number of sections.
As for analog VLSI, fitting several hundred segments on a
single chip is possible if small transistors and capacitors are
used, but these are prone to mismatch and noise. However, given
that the biological cochlea itself is built out of imprecise components, we conjectured that ABC will be robust to mismatch
and noise in Si devices. For instance, noise in transistors used to
model the fluid in the Si cochlea parallels Brownian motion of
water molecules impinging on the basilar membrane. Our motivation for implementing ABC in analog VLSI was to explore
this conjecture—if indeed ABC inherited its biological counterpart’s robustness. To this end, we integrated hundreds of BM
segments in a single chip, passively and actively coupled by
transistors mimicking the cochlear-fluid and ABC, respectively.
In addition, the chip includes Si neurons that convert analog signals, representing BM velocity, into digital pulse-streams, representing AN fibers’ spike trains.
Fig. 1. Cochlea. (a) Cutaway showing cochlear ducts (adapted from [20]), comprising (inset) the scala vestibuli (SV), scala media (SM), and scala tympani
(ST). The cochlear partition (CP) separates the two perilymphatic scalae. (b)
Longitudinal view of the CP (adapted from [7], [21]–[23]). Outer hair cells
(OHCS) tilt toward the base while Deiters’ cells’ (DC) phalangeal processes
(PhP) tilt toward the apex; their bases rest on the basilar membrane (BM) and
their tips form the reticular lamina (RL). d is the tilt distance.
in thickness, resulting in an exponential decrease in stiffness,
which gives rise to the passive frequency tuning of the cochlea.
BM vibration is actively enhanced by OHC electromotile
forces, resulting in the cochlea’s exquisite sound sensitivity,
frequency discriminability, and nonlinearity.
Assuming it is incompressible, the fluid’s motion can be described by a velocity potential that satisfies
,
where
is the Laplacian operator; is the distance from the
stapes along the BM with
at the base (or the stapes); and
is the vertical distance from the BM, with
0 at the BM.
By definition, the velocity potential is related to the fluid velocity’s components in the and directions:
and
[5].
The BM’s response to both the pressure difference ( ) between the fluid ducts3 and the OHC forces (
) can be described as
(1)
where
,
, and
are, respectively, the BM’s stiffness, damping, and mass (per unit area) and is the BM’s downward displacement. The pressure difference is given by
, evaluated at the
BM (
0); is the fluid density.
The
term combines forward and backward OHC forces
[Fig. 1(b)], described as in [8]
(2)
III. NONLINEAR ACTIVE COCHLEAR MODEL
The cochlea actively amplifies acoustic signals as it performs
spectral analysis. Incoming sound moves the oval window
(stapes) at the cochlea’s base, which, in turn, sets the cochlear
fluid in motion [Fig. 1(a)]. The fluid interacts with the BM,
the cochlea’s main vibrating organ, forming a traveling wave
that propagates toward the cochlea’s apex. From the base to
the apex, BM transverse fibers increase in width and decrease
2The
coupling in [18] is between automatic-gain-control (AGC) filters, not
between BM segments.
where represents OHC motility, expressed as a fraction of BM
stiffness, and is the ratio of forward to backward coupling,
representing relative strengths of OHC forces exerted on the BM
segment directly through a Deiters’ cell (DC) on which the OHC
sits (first term), and indirectly via a phalangeal process (PhP) attached to the reticular lamina (RL) (second term).
and
3Only the scala vestibuli and the scala tympani are considered, since
Reissner’s membrane, which separates the scala vestibuli and the scala media,
is extremely thin, presenting negligible acoustic impedance [23].
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
446
IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 6, DECEMBER 2009
are the displacements of adjacent upstream and downand
), respectively; denotes
stream BM segments (
the tilt distance, the horizontal displacement between the source
and the recipient of the OHC force, assumed to be equal for the
forward and backward cases. The function models saturation
of OHC forces, a nonlinearity evident in physiological measurements [23].
The forward and backward coupling forces’ opposite signs
account for the fact that OHCs move the BM and the RL in opposite directions. Forward coupling, proposed by others [24],
’s
[25], posits that the OHC’s basal tilt results in Segment
BM motion reinforcing that of Segment . Backward coupling,
the novel component of ABC, posits that the PhP’s apical tilt
’s motion opposing that of Segment .
results in Segment
Adding ABC to a passive model makes the peak in BM displacement higher and sharper, similar to the difference between a live
and dead cochlea. This frequency-selective amplification arises
because ABC makes the damping negative when the wavelength
becomes short (see [8] and [17] for further details).
IV. CIRCUIT DESIGN
Based on the mathematical cochlear model, we design a 2-D
nonlinear active cochlear circuit in analog VLSI, taking advantage of the 2-D nature of Si chips. We start by synthesizing a
passive model, and then extend it to a nonlinear active one by
including ABC with saturation.
decomposition that provides economy and accommodates ABC
readily (see Section IV-C)
(4)
The LPFs’ outputs are and
(a.k.a., state variables); their
time constants are and , respectively; is a gain factor.
The BM’s velocity matches the fluid’s; thus, we must ensure
that
(recall that, by definition,
). The RHS is proportional to the current in the diffusor that connects these two
nodes. Therefore, we can satisfy this constraint simply by con) to the fluid cirnecting the BM circuits’ current output (
cuit—and setting the diffusors’ gate voltage (
; see Fig. 7)
appropriately (i.e.,
, where is the nMOS transistors’ subthreshold slope coefficient and
is the thermal
voltage, 25.6 mV at room temperature).
B. Circuit Analogs of Biology
Given (4), ,
output current
, and
can be expressed in terms of the
A. Passive Cochlear Circuit
The model consists of two fundamental parts: 1) the cochlear
fluid and 2) the BM. First, we design the fluid circuit by using the
discrete version of Laplace’s Equation (in 1-D for simplicity)
(5)
where
. The velocity potential may be represented in one of two ways: If node ’s voltage represents
(voltage-mode), resistors connect adjacent nodes. If node ’s
voltage represents the
(log domain), subthreshold MOS
transistors (diffusors) connect adjacent nodes [26], [27]. The
latter is simpler to implement (see Fig. 7): We used nMOS transistors for the diffusors and pMOS transistors to take the antilog,
yielding a current that is proportional to , a good approximation if the pMOS transistors’ (subthreshold slope-coefficient) is close to one.
Second, we design a BM segment and, thus, the BM. If
represents
(the velocity potential scaled by the fluid density) and
represents (BM velocity), the BM boundary
condition (1) can be expressed as
By comparing the expression for
we obtain the circuit counterparts
with the design target (3),
and
(6)
where the mass is normalized. These analogies require that the
time constants ( and ) increase exponentially to simulate the
exponentially decreasing BM stiffness (and damping). allows
us to achieve a larger quality factor (a measure of frequency
selectivity) for a given choice of and (limited by capacitor
size :
, where is the current level). That is
These circuit-biology relationships help determine the parameter values used in circuit simulation and chip operation.
C. Adding Active Bidirectional Coupling
(The
term is dealt with in Section IV-C.) Taking the first
time derivative and working in the -domain (
) yields
We synthesize an active BM segment by following the same
procedure we used for the passive one, but with the
term
included. The design target equation becomes
(3)
We synthesized this second-order system from two low-pass filters (LPFs) by using a custom Mathematica program to find a
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
WEN AND BOAHEN: SILICON COCHLEA WITH ACTIVE COUPLING
We find
state variable
447
(i.e., the time-integral) by observing that the
in the passive design (5) is related to
by
. Thus
where
and
represent the output currents, and
and
are the time constants of the first LPF in the upstream and downstream BM segment, respectively. We replace
and
by
—the receiving segment’s time constant—a good approximation due to the small change in between neighboring segments.
Therefore, the design target becomes
(7)
[see (6)]— was factored
where
out by rescaling ;
and
denote the forward
and backward OHC force factors, respectively.
We synthesized the circuit following a procedure similar
to that used in the passive design. Only the second equation
changed
(8)
and
. Note that to
where
include ABC, we need to only add two currents to the input
of the second LPF in each BM segment circuit; these currents
are from its adjacent neighbors. Specifically,
and
are the
output currents ( ) of the first LPF in the upstream (basal) and
downstream (apical) BM segments, respectively.
Fig. 2. Low-pass filter (LPF) and basilar membrane (BM) segment circuits. (a)
Half-lowpass—filter circuit (* denotes complex current mirror). (b) Complete
LPF circuit formed by two half-LPF circuits. (c) BM-segment circuit. It consists
of two LPFs and connects to its neighbors, sending current I , and receiving
current I (last two terms in (8), corresponding to ABC). In (b) and (c), current
splitting implies current copying.
V. CIRCUIT IMPLEMENTATION
Based on our synthesized design, we implement a Class AB
log-domain circuit for the BM segment. We employ the log-domain filtering technique [28] to realize current-mode operation.
In addition, we adopt Class AB operation to increase dynamic
range, reduce the effect of mismatch, and lower power consumption ([29]–[31]). This differential signaling is inspired by
the biological cochlea—the BM’s displacement is driven by
the pressure difference across it. We present the transistor-level
schematics in this section as well as an Si neuron that converts
the segment’s output current into a pulse stream.
A. Basilar Membrane Circuit
Taking a bottom-up approach, we start by designing a Class
AB LPF, a building block for the BM circuit. An LPF is described by
where
is the input current,
is the output current, and
sets the time constant. Its differential counterpart is
where each signal is expressed as the difference between its
positive ( ) and negative ( ) components. The common-mode
constraint is
where
sets the geometric mean of the output current’s
components.
Combining the common-mode constraint with the differential
design equation yields the positive path’s nodal equation (the
negative path has superscripts and swapped) [29]
This nodal equation suggests the half-LPF circuit shown in Fig.
2(a).
, the voltage on the positive capacitor ( ), gates a
pMOS transistor to produce the corresponding current signal
(
and
are similarly related). The bias
sets the
quiescent current while
determines the current , which
is related to the time constant by
. Two of these
subcircuits, connected in push-pull, form a complete LPF [Fig.
2(b)]. Specifically, when the input
charges
, it also dis; similarly,
charges
and discharges
.
charges
The BM-segment circuit [Fig. 2(c)] is implemented by using
two LPFs interacting in accordance with the synthesized design equations.
is the sum of three signals , , and
(4). The positive and negative components of ,
, and
are the differential output currents of the first LPF (with
time-constant ), corresponding to
and
in the LPF
symbol [see Fig. 2(b)], respectively. Similarly,
and
are
the output currents of the second LPF (with time-constant ).
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
448
IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 6, DECEMBER 2009
Fig. 3. Summing, scaling, and subtracting circuits. Left: Summing circuit.
Tying three wires together sums their currents. Middle: Scaling circuit. Setting
V
V
b = scales the current (I ) made by a pMOS transistor
with the same gate voltage (V ) by b. Right: Subtracting circuit. Tying two
wires to either side of a complex current mirror (*) subtracts their currents;
passing the result through a diode-connected transistor rectifies it. Superscript
p corresponds to I
I , while m corresponds to I
I (see Fig. 4).
=
+ u ln( )
0
0
Fig. 5. Inner hair cell circuit. It converts differential currents, representing BM
velocity, into a low-pass filtered, half-wave rectified current that drives the spiral
ganglion cell (SGC) circuits. V sets the quiescent level, V sets the current
sets the synaptic
splitter’s gain, V sets the low-pass filter’s time-constant, V
efficacy, and V sets the maximum amplitude.
Fig. 4. Active-coupling circuitry. BM Segment i receives a scaled and saturated
version of Segment i
and i
’s I outputs. The circuitry that couples
Segment i to Segments i
and i
is omitted for clarity. V and V set
the gain of feedforward and feedbackward coupling, respectively; V sets their
saturation levels.
+1
01
01
+1
Summing ( ) is implemented by exploiting Kirchhoff’s Current Law (Fig. 3); scaling ( ) is implemented by biasing a pMOS
transistor’s source voltage (Fig. 3).4
ABC is implemented by exchanging currents between neighand reboring BM segments (Fig. 4). Each BM sends out
ceives , a saturated and scaled version of its neighbor’s (8).
The saturation is accomplished by a current-limiting transistor,
which yields
[32], where
is set by a bias voltage . We used a subtract circuit (Fig. 3) to
take the difference first because saturation is applied to the differential signal, not to its positive and negative components. The
scaling corresponds to the gain factors
and
in (8), implemented by biasing a pMOS transistor’s source voltage ( and
in Fig. 4(b), respectively).
B. Spiral Ganglion Cell Circuit
In the biological cochlea, the BM’s velocity, sensed by inner
hair cells (IHCs), is encoded by spiral ganglion cells (SGCs).
Behaving like pulse-frequency modulators, SGCs convey information about sound stimuli—including frequency, level, and
timing—over the AN (axons of SGCs). Their spikes are evoked
by neurotransmitter released from IHCs, each of which drives
10 to 30 SGCs, increasing from apex to base [33]. In our Si
cochlea, this fanout is 6.
The IHC circuit has three functional components (Fig. 5): 1) a
current mirror takes the difference between
’s positive and
negative components; 2) a current splitter half-wave rectifies the
4In the case of I
, the sum is mirrored twice to produce additional copies
to feed to the scanner and pulse-frequency modulators.
Fig. 6. Spiral ganglion cell circuit. It models three membrane-voltage-dependent currents (I , I , and I ) and one Ca-concentration-dependent current
(I
). These currents turn on when V exceeds certain levels (set by V
and V
) or when a spike occurs (activates Req, which activates Ack). I
sets I ’s time-constant; V and I
set the Ca-concentration’s increment
(per spike) and time constant, respectively.
difference, and 3) a class A log-domain LPF filters the halfand
) can
wave-rectified currents. The bias voltages (
be varied to yield distinct rate-level relations (i.e., sound level
to spike rate).
Augmenting its static rate-level relation, an SGC’s dynamic
properties enhance the encoding of a sound stimulus’ temporal
features: It fires at a higher rate at stimulus onset, due to the
presence of a Ca-concentration-dependent K-current [34]. And,
from cycle to cycle, it is more likely to fire when the sinusoid
is rising most rapidly (phase locking [35]), due to the presence
of a low-threshold K-current [36], [37]. In addition to these two
K-currents, the SGC circuit models an Na current that generates
an all-or-none spike and a high-threshold K-current that resets
the membrane [38] (Fig. 6).
An address-event encoder transmits the SGC circuits’ spikes
off-chip [39]–[41]. To communicate with the address-event
encoder, the SGC makes a request when it spikes and clears
this signal when acknowledged (see Req and Ack in Fig. 6).
The spike is encoded as a unique address (specifying row and
column). The receiver chip decodes this address and delivers
the spike to the target neuron.
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
WEN AND BOAHEN: SILICON COCHLEA WITH ACTIVE COUPLING
Fig. 7. Active-coupling architecture. Differential audio signals are applied at
the base of two diffusive-element grids (representing the top and bottom fluid,
respectively; V determines the fluids’ density), connected at the apex by a
single diffusive element (representing the helicotrema). Second-order sections
(representing the BM segments) embedded between the grids send/receive
currents to/from their immediate neighbors (V as in Fig. 4), realizing active
, representing BM
bidirectional coupling (ABC). Their output currents (I
velocity) each drive six pulse-frequency-modulation circuits (representing the
spiral ganglion cells (SGCs)).
C. Chip Architecture
We fabricated a version of this design with 360 BM-segment
circuits, two 4680-element (360 13) fluid grids, and 2160
(360 6) SGC circuits (Fig. 7). The number of BM segments
was chosen to satisfy the requirements of our software simulation—the chip has approximately 55 segments per octave
(assuming a 200–20 kHz range, or 6.6 octaves). The fluid-element grids’ height (13) was chosen to match the biological
cochlea’s aspect ratio, a factor important in controlling the
traveling wave’s behavior [5]. The number of SGC circuits per
BM segment (6) was chosen to ensure that the stimulus evokes
multiple spikes per cycle—an octave-wide response will produce up to 33 kspikes/s (assuming a maximum spike frequency
of 100 Hz). A die photo of the chip is shown in Fig. 8.
VI. CHIP RESPONSES
We measured the Si cochlea’s BM responses to pure tones and
its AN responses to complex sounds. To supply sinusoidal current as input, we applied the logarithm of a half-wave-rectified
sinusoid to the top and bottom fluid grids; these two voltage signals were 180 out-of-phase. Scaled to match the pMOS transistors’ subthreshold slope-coefficient (
0.58, measured), the
half-wave rectified voltage signals’ peak amplitude varied from
0.12 to 0.36 V, dropping from a baseline of 2.26 V (
2.4
V), in 0.04-V steps. These values correspond to input current
amplitudes of 0 to 48 dB, increasing in 8-dB steps, with 24 dB
corresponding to a medium sound level. We set up the cochlea
chip’s time-constant-setting voltages (
and
) by tuning
449
the base and the apex to approximately 20 kHz and slightly
below 200 Hz, respectively. Linear interpolation (implemented
with two polysilicon lines spanning the Si cochlea’s length)
gave rise to exponentially decreasing time-constant currents.
The saturation level of ABC currents was set to its maximum
1.7 V, putting the pMOS transistor above threshold)
level (
unless otherwise stated.
We measured frequency responses as well as longitudinal
responses. To obtain frequency responses, we swept the input
frequency and measured BM current outputs (from both positive and negative paths) at a particular segment.5 To obtain
longitudinal responses, we kept the input frequency fixed and
measured current outputs at consecutive segments along the
cochlea’s length. Selecting a particular segment or sweeping
through consecutive ones is realized with a built-in scanner
(modified from [42] to accommodate snaking). AN responses,
on the other hand, were measured in parallel by capturing
(time-stamped) address-events over a universal-serial-bus
(USB) link.
Here, we present frequency responses measured from linearly spaced BM segments and longitudinal responses to octave-spaced pure tones, both at an input level of 24 dB. We
also map the dependence of frequency and signal-to-noise ratio
(SNR) on position, also at a 24-dB input level. In addition, we
present frequency responses obtained at various input intensities (0 to 48 dB), demonstrating automatic gain control. We then
demonstrate the role of ABC by disabling it. Finally, we present
the chip’s real-time responses to a chirp-click sound sequence.
A. Frequency Responses
Frequency responses reveal the tuning of individual BM segments (Fig. 9).6 Despite some irregularities in response shape
and peak height (due to transistor mismatch), the chip’s responses captured the characteristics of the biological responses,
at least qualitatively. Frequency responses are peaked and cutoff
slope is steep (more so in some segments than others), with peak
or characteristic frequencies (CFs) ranging from 13.8 k to 218
Hz for these six BM segments (40-segment spacing). Phase accumulates gradually at first, then more rapidly near the peak
[marked by dots in Fig. 9(b)], and plateaus after the peak. The
large accumulation indicates a traveling wave; the plateau indicates its extinction.
Histograms of measurements from 12 equally spaced BM
segments reveal marked differences across the cochlea, indicating poor parameter matching among segments, to the
extent that desired performance was not achieved at all taps
(Fig. 10). Tip-to-tail ratios (amplitude difference between the
peak and lowest frequency point), a commonly used measure
of cochlear amplification, ranges from 9 to 45 dB, approaching
the chinchilla’s performance (53 dB). CF phase ranges from
0.4 to
radians, spanning the chinchilla’s performance
(
).
ranges from 0.1 to 2.7, reaching the
chinchilla’s performance (2.55 at medium sound levels). The
cutoff slope ranges from
to
dB/octave, falling 1.6 times
or more short of the chinchilla (
dB/octave). In addition,
5Segment
numbers increase from base to apex, starting from 1.
beyond 240 were not considered because they did not respond
robustly, probably due to the large discontinuity we observed at each U-turn
(Segments 60, 120, 180, etc.), presumably caused by doping-level deviations at
the array’s edges (dummy cells were not deployed).
6Segments
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
450
IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 6, DECEMBER 2009
Fig. 8. Die photo. Fabricated in a 5M 1P 0.25-m CMOS process, the ABC
cochlea, with six 60-segment columns snaking to yield a desirable aspect ratio,
occupies 10.9 mm . Input, basilar membrane, top/bottom fluid, auditory nerve
(axon of spiral ganglion cells, or SGCs), scanner, and address-event encoder
circuits are labeled.
Fig. 10. Histograms of measurements from 12 BM segments, spaced 20 segments apart, from 10 to 230 (24-dB input level). (a) Tip-to-tail ratio. (b) Peak
phase. (c) Q . (d) Cutoff slope.
Fig. 11. Frequency-position map and signal-to-noise ratios (SNRs) at the
24-dB input level. (a) The frequency a segment responds maximally to (CF;
dots) is logarithmically related to its position (line). (b) SNR at 12 cochlear
segments (for CF). Dots: Data; line: Linear regression.
segments, stimulated at their CF [Fig. 11(b)]. SNR was computed as the ratio between the signal’s power (i.e., squared amplitude at the CF) and the noise’s power (sum of squared amplitude at all frequencies—see Fig. 14(a)). A linear regression of
SNR versus cochlear position yielded
, indicating insignificant accumulation.
B. Longitudinal Responses
Fig. 9. Frequency responses of six BM segments, spaced 40 segments apart,
from 30 to 230 (24-dB input level). (a) Amplitude. (b) Phase (dots mark the
characteristic frequencies). Biological data are provided for comparison (dashed
line, chinchilla measurement at the medium sound level [43]).
similar to the chinchilla cochlea’s basal region, the Si cochlea
has a logarithmic frequency-position map: Segment number
( ) is related to CF ( , in Hertz) by
[Fig. 11(a)].
To evaluate noise accumulation in our ABC architecture, we
calculated the SNR at each of the 12 equally spaced cochlear
Longitudinal responses give a snapshot of the entire basilar
membrane, thereby providing a direct measurement of the traveling wave, whose wavelength ABC is sensitive to. They also
show how the wave’s amplitude builds up as it travels from the
base to the apex, providing evidence that ABC acts in a distributed fashion. The chip’s longitudinal responses show large
variations from segment to segment (due to mismatch), which
we filtered with a 10-segment moving average in order to estimate the response characteristics [Fig. 12(a)].
We measured longitudinal responses to four pure tones, with
octave spacing [Fig. 12(b)]. A 4-kHz tone elicits a peak response
at Segment 85 (characteristic place, or CP) while a 500-Hz tone
travels further and peaks at Segment 178. The CPs for the two
intermediate frequencies (1 and 2 kHz) are Segment 166 and
139, respectively. Tip-to-tail ratios range from 12 to 32 dB;
s
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
WEN AND BOAHEN: SILICON COCHLEA WITH ACTIVE COUPLING
451
Fig. 14. Signal-to-noise ratio (Segment 100). (a) Spectra of output to near-CF
input (5-kHz tone input at 40 dB). Interpolation (unfilled dots) was used to estimate noise at the input frequency and its harmonics. (b) Output SNR increased
with increasing input amplitude at low intensities but saturated above 24 dB.
Fit: Michaelis–Menten function.
range from 0.9 to 1.2; and cutoff slopes range from
dB/octave.7
16 to
70
C. Input/Output Functions
Fig. 12. Longitudinal responses. (a) Raw and smoothed longitudinal responses
(2 kHz tone input at 24 dB). A 10-segment moving average removes the large
segment-to-segment variations. (b) Smoothed longitudinal responses to four octave-space frequencies. As frequency increases (from 500 to 4 k Hz), the response peaks closer to the Si cochlea’s base.
Fig. 13. Nonlinear compression. BM-velocity frequency responses for different input amplitudes (Segment 100; CF = 5.6 kHz). (a) Amplitude. Equally
spaced responses indicate linear behavior. (b) Phase. Inset: Input–output functions, measured at CF, an octave below, and half an octave above. Biological
measurement is provided for comparison (open circles, chinchilla measurement
from [43], shifted to align the lowest input level tested with that of the chip).
Dotted line: Identity (y = x).
Input/output (I/O) functions reveal the Si cochlea’s nonlinear
behavior. We increased the input amplitude exponentially
by increasing the voltage applied linearly, calculating the
corresponding amplitude (in decibels) based on the chip’s
subthreshold slope-coefficient (measured experimentally). The
amplitude range applied was constrained on the high side by
strong inversion (leaving the subthreshold region), and on the
1.9 V to saturate active
low side by the noise floor. We set
coupling at the upper end of this range, thereby producing
compression.
BM responses show compressive growth, first at the CF and
then at nearby frequencies (Fig. 13). As a result, BM responses
become more broadly tuned with increasing input amplitude;
drops from 1.8 to 1.1. There is a corresponding decrease in
cutoff slope, which drops from 44 to 13 dB/octave. Unlike
biology, where there is a basal shift (to lower frequency) [44],
the CF hardly changes, probably due to insufficiently high input
levels. Response phase does not change significantly; this is the
case in biology as well. The larger phase plateaus (exactly
apart) at low input amplitudes (0 and 8 dB) are due to noisy
responses in the cutoff region.
Compression does not occur symmetrically around the peak:
It sets in at lower intensities for frequencies below the CF (see
Fig. 13, inset). Whereas at the CF (5.6 kHz), compression sets in
when the input amplitude exceeds 24 dB, one octave below (2.8
kHz) it occurs at 32 dB, and half an octave above (7.9 kHz),
it occurs at 48 dB (the largest amplitude applied). This result
suggests that upstream segments (higher CFs) contribute to automatic gain control, more so than downstream segments (lower
CFs).
The chip’s CF behavior agrees qualitatively with the chinchilla measurements (see Fig. 13, inset), except that at high intensities, which the chip input did not reach, the chinchilla’s
I/O function became linear again, resembling a passive cochlea.
Above or below the CF, the chip’s I/O functions are less linear
(more compressive) than the chinchilla’s (data not shown), presumably because the chip’s tuning is broader so that compression at high sound levels occurs with a larger spread.
To find the lowest detectable input amplitude, we measured
SNR at the output (defined as the signal-squared over noise7The number of segments spanned by an octave was calculated from the CF
range of the first 240 segments.
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
452
IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 6, DECEMBER 2009
Fig. 16. Longitudinal responses with (active; black) and without (passive;
gray) coupling (2-kHz tone input at 24 dB).
Fig. 15. Frequency responses for various coupling saturation levels (Segment
100; 16-dB input level). (a) Amplitude. (b) Phase. The saturation level decreases
as V increases.
squared) for various input amplitudes and extrapolated to 0 dB
(Fig. 14). Output SNR increased from 2.8 to 19 dB as input amplitude increased from 0 to 48 dB. A Michaelis–Menten function fitted the SNR’s initial increase and asymptotic behavior
well:
, with
64.3,
5.9,
and
1.8, where and represent the SNR and input amplitude (relative to the smallest current applied), respectively.8
Extrapolating the fit yields an output SNR of 1 (i.e., 0 dB) at
an input amplitude of 4 dB, indicating a 52-dB input dynamic
range.
D. Effect of Active Bidirectional Coupling
Varying the coupling’s saturation level (through ) demonstrates the ABC’s role. In all responses presented thus far, except for Section VI-C, the saturation level was high enough to
avoid saturation. It gets progressively lower as
(which gates
pMOS transistors) increases, producing saturation at lower and
lower input levels. Coupling is negligible for
2.2 V, which
corresponds to a passive cochlea.
We obtained a series of frequency responses from Segment
100 with different saturation levels (Fig. 15). Decreasing saturation levels resulted in smaller response amplitudes. The amplitude decreased monotonically from 33.3 to 15.1 dB (arbitrary
scale) at the CF, an 18 dB drop. Since decreases were more
prominent in this region, responses became more broadly tuned;
decreased monotonically from 1.14 to 0.45. The phase did
not change significantly—except for the weakest couplings.
We also measured longitudinal responses to a 2-kHz tone with
(
0.26 V) and without (
2.35 V) coupling (Fig. 16).
The peak amplitude was 14.6 dB larger with coupling; the cutoff
increased from 0.39 to
slope was 35 dB/octave steeper;
1.16. These increases are comparable to those seen in Segment
100’s frequency response (increases of 18.2 dB, 22.0 dB/octave,
and 0.69, respectively).
8To
convert to decibels, take log
spectively.
of
x or R and multiply by 20 or 10, re-
In summary, ABC increases gain and sharpens tuning,
achieving responses that are qualitatively comparable to physiological measurements. Indeed, the cases with and without
ABC resemble live (active) and dead (passive) cochlea, respectively; thus, ABC captures the role of OHC electromotility—at
least qualitatively.
E. Si Auditory Nerve
We visualized the Si AN’s response by constructing a
cochleagram (Fig. 17). This raster plot displays spike trains
of all SGCs in Segments 1 to 240, a total of 1440 (240 6)
outputs,9 with time running from left to right and the segment
number running from top (base) to bottom (apex)—high to low
frequency.
The Si AN responds to the chirp-click sequence with a wave
of spike activity followed by a flash (see Fig. 17). The wave
propagates from the base to the apex in response to the chirp’s
decreasing frequency. It becomes more sharply defined after
the first 60 segments (i.e., 360 SGCs), indicating the extent
to which frequency selectivity arises cooperatively. The flash
lights up all outputs simultaneously in response to the click’s
broad frequency content, except for apex, where it is masked by
the chirp’s close proximity in time (contiguity). This masking is
due to SGC spike-rate adaptation, which emphasizes sound onsets. A few highly excitable SGCs (e.g., Channel 223 and 480)
respond throughout most of the stimulus; this behavior is due to
transistor mismatch.
In summary, the Si AN encodes a sound’s frequency, intensity, and timing. It uses a place code for frequency: only neurons
at a certain location fire. It uses a rate code for intensity: these
neurons spike at higher rates. And it uses a real-time code for
timing: spike rates change in real time, with sound onset emphasized by SGC spike-rate adaptation. The chip’s specifications
are summarized in Table I.10
VII. DISCUSSION
The chip measurements presented here demonstrate that ABC
overcomes the major shortcomings of previous Si cochlea architectures, summarized in Table II. In the cascade architecture,
noise increased a hundredfold (asymptoting after 30 segments)
[3]. With ABC, noise does not accumulate, as demonstrated by
our SNR measurements [Fig. 11(b)]. In the passively coupled
9SGCs from Segment 241 to 360 (apical third of the cochlea) were omitted
for the same reason stated earlier.
10Mean
6 standard deviation is quoted for n
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
=
12 measurements.
WEN AND BOAHEN: SILICON COCHLEA WITH ACTIVE COUPLING
453
TABLE III
Si COCHLEA DEVICE SIZES
Fig. 17. Chirp-click cochleagram. The chirp invokes a wave that propagates
from the base (top) toward the apex (bottom). The click invokes a flash that
lights up all but the lowest frequency outputs. Inset: Chirp-click sound. On a
logarithmic scale, the 1.5-s chirp’s frequency decreases linearly and the 0.1-s
click’s 50 discrete frequencies are equally spaced. Both span 16 kHz to 200 Hz.
TABLE I
Si COCHLEA SPECIFICATIONS
TABLE II
Si COCHLEA ARCHITECTURES
architecture, gain decreased by 25 dB [6]. With ABC, this destructive interference is overcome, as demonstrated by our gain
and tuning measurements [Figs. 15 and 16]. However, ABC’s
gain increase was limited to 18 dB by mismatch-induced traveling-wave reflections. We confirmed that these reflections can
reduce gain and broaden tuning by performing simulations with
mismatch included.
The dominant source of transistor mismatch is thresholdvoltage variation, which has been shown to be Gaussian distributed, with variance inversely proportional to the transistor’s
channel area [47]. When transistors operate in weak inversion
for low power consumption, their currents are log normally
distributed. For instance, in a 0.35- m CMOS process, the currents’ coefficient of variation (standard deviation over mean),
or CV, is 9.2% and 22% for 11.4
11.4 and 4.6
4.6
(
0.18 m) nMOS transistors, respectively [48]. Given the
transistors sizes in our circuits (Table III), we used log normally
distributed parameter values with CVs ranging up to 25% in
our simulations (see the Appendix).
We quantified the mismatch’s effect on active amplification
), extracted from
(tip-to-tail ratio) and tuning sharpness (
smoothed BM velocity responses (Fig. 18). With OHC motility
factor
0.15, these metrics dropped from 85 4 (mean
standard deviation) to
7 dB and from 5.2 0.6 to 1.1
0.8, respectively, as CV increased from 5% to 25%, becoming
similar to the passive case (
0), albeit with substantially
larger variance. This loss of sensitivity and selectivity could be
counteracted by increasing . For
25%, increasing
from 0.15 to 0.25 increased peak gain from 41 7 to 70 13
dB and
from 1.1 0.8 to 3.6 1.5, with the variance increasing dramatically in both cases. For comparison, with
0.15, these metrics were 89 dB and 5.6, respectively, in the absence of mismatch.
These simulation results suggest that mismatch accounts
for shortfalls in the chip’s overall performance as well as
variability among its segments. For the value of the chip used
(0.14—estimated from bias voltages that determine , , and
), the simulations reproduce the range of values we measured
for tip-to-tail ratio and
(see Table I) with a parameter
CV of slightly above 20% [see Fig. 18(c) and (d)], which is
twice the 10% current CV of the chip’s (mostly) 10
10
transistors.11 We could not confirm the predicted performance
improvement with values of greater than 0.14 because they
produced instability in the chip.
These simulation results also suggest that ABC has the potential to exceed the best performance achieved by Si cochleae
to date. Sarpeshkar et al.’s hybrid parallel-cascade architecture
achieved a tip-to-tail ratio of 77 dB [3]. Whereas, Fragniére’s
passively coupled architecture achieved a
of 2.34 [6]. In
comparison, our simulations predict that with the practical OHC
motility factor of 0.14, ABC can achieve a tip-to-tail ratio of
of 5.2 0.6 if parameter-CV is reduced
85 4 dB and a
to 10%. This requires decreasing the current CV from 10% to
5% by increasing transistor sizes from 10 10 to 20 20 .
These larger transistors would only increase the chip’s size from
10.9 to 14.3 mm , since the BM segments currently occupy
only 10.3% of its area (excluding the capacitors). However, this
promised performance remains to be proven in Si.
VIII. CONCLUSION
We presented a mixed-signal VLSI implementation of a 2-D
nonlinear cochlear model that utilizes a novel cochlear amplifier mechanism, ABC. ABC produces large amplification and
sharp tuning to soft sound and nonlinear compression and broad
11This doubling could be produced by cascading two current mirrors. The fact
that this is less than the actual number of mirroring operations in a BM segment
is explained by the averaging that occurs when several segments’ outputs are
actively and passively coupled together to yield the measured response.
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
454
IEEE TRANSACTIONS ON BIOMEDICAL CIRCUITS AND SYSTEMS, VOL. 3, NO. 6, DECEMBER 2009
REFERENCES
Fig. 18. Simulated responses (2-kHz tone). Parameter coefficient-of-variation
(CV) and OHC motility factor () were varied. (a) and (b) BM velocity’s lon0.15 and
10% and 20%. A ten-segment
gitudinal response with moving average yielded smoothed responses. (c) and (d) Tip-to-tail ratio and
Q decrease with CV. Datapoints for 0 and 0.25 were shifted slightly
to avoid overlap with those for 0.15; error bars give 1 standard deviation
for 100 trials.
=
=
CV =
=
=
6
tuning to loud sound. Rather than detecting signal amplitude
and implementing an AQC loop, ABC simply mimics OHCforce saturation, which acts instantaneously; and ABC senses
the wavelength, which adds frequency selectivity. It successfully combats destructive interference in the coupled architecture.
ABC was vulnerable to reflections that occur when BM
properties change abruptly due to device mismatch. However,
reducing mismatch by increasing transistor area will decrease
the number of BM segments per octave, which will broaden
frequency tuning. Alternatively, as mismatch perturbs BM
properties at the finest spatial scale, it could be counteracted by
increasing the wavelength, something the three-to-five-OHC
tilt observed in the Organ of Corti [21], [22] would achieve.
The caveat is that lengthening the wavelength will also broaden
tuning. Further study is required to determine which approach
offers the most favorable tradeoff.
APPENDIX
Here, we describe our simulation procedures. We formulated
a linear version (no saturation) of the model described in Section III using circuit variables (i.e., , , , and
) and
parameters (i.e., and ), based on (4) and (8). We discretized
the model horizontally and vertically into a 360 13 grid (identical to the chip). We obtained solutions for the fluid’s velocity
potential ( ) and the BM’s velocity (
) at each location
using the finite difference method [49] in the frequency domain.
On each trial, we introduced log normally distributed variations
to and . We assumed they were identically distributed for
simplicity (i.e., equal mean and variance). The mean was determined based on the biophysical analogs (6). We also applied log
normally distributed variation to the cochlear fluid’s density ( ).
ACKNOWLEDGMENT
The authors would like to thank J. V. Arthur for his help with
debugging the chip-PC USB interface used to visualize spike activity. They would also like to thank the Associate Editor T. Delbrück and three anonymous reviewers for their valuable comments and suggestions on improving this presentation.
[1] J.-J. Sit and R. Sarpeshkar, “A cochlear-implant processor for encoding
music and lowering stimulation power,” IEEE Pervasive Comput., vol.
1, no. 7, pp. 40–48, Jan.–Mar. 2008.
[2] R. F. Lyon and C. A. Mead, “An analog electronic cochlea,” IEEE
Trans. Acoust. Speech Signal Process., vol. 36, no. 7, pp. 1119–1134,
Jul. 1988.
[3] R. Sarpeshkar, R. F. Lyon, and C. Mead, “A low-power wide-dynamicrange analog VLSI cochlea,” J. Analog Integr. Circuits Signal Process.,
vol. 16, no. 3, pp. 245–274, 1998.
[4] R. Sarpeshkar, C. Salthouse, J.-J. Sit, M. W. Baker, S. M. Zhak, T. K.-T.
Lu, L. Turicchia, and S. Balster, “An ultra-low-power programmable
analog bionic ear processor,” IEEE Trans. Biomed. Eng., vol. 52, no. 4,
pp. 711–727, Apr. 2005.
[5] L. Watts, “Cochlear mechanics: Analysis and analog VLSI,” Ph.D. dissertation, Calif. Inst. Technol., Pasadena, CA, 1993.
[6] E. Fragniére, “A 100-channel analog CMOS auditory filter bank for
speech recognition,” in Proc. IEEE Int. Solid-State Circuits Conf. Dig.
Tech. Papers, 2005, pp. 140–141.
[7] W. E. Brownell, C. R. Bader, D. Bertrand, and Y. Deribaupierre,
“Evoked mechanical responses of isolated cochlear outer hair cells,”
Science, vol. 227, no. 4683, pp. 194–196, 1985.
[8] B. Wen and K. A. Boahen, “A linear cochlear model with active bi-directional coupling,” in Proc. 25th Annu. Int. Conf. IEEE Eng. Med.
Biol. Soc., 2003, pp. 2013–2016.
[9] B. Wen and K. A. Boahen, “A 360-channel speech preprocessor that
emulates the cochlear amplifier,” in Proc. IEEE Int. Solid-State Circuits
Conf. Dig. Tech. Papers, 2006, pp. 556–557.
[10] S. Martignoli, J.-J. van der Vyver, A. Kern, Y. Uwate, and R. Stoop,
“Analog electronic cochlea with mammalian hearing characteristics,”
Appl. Phys. Lett., vol. 91, no. 6, pp. 064108–064108, 2007.
[11] T. Hamilton, C. Jin, J. Tapson, and A. van Schaik, “A 2-D cochlea
with Hopf oscillators,” in Proc. IEEE Biomedical Circuits and Systems
Conf., 2007, pp. 91–94.
[12] R. Sarpeshkar, R. F. Lyon, and C. Mead, “An analog VLSI cochlea
with new transconductance amplifiers and nonlinear gain control,” in
Proc. IEEE Int. Symp. Circuits and Systems, Atlanta, GA, 1996, vol. 3,
pp. 292–295.
[13] T. J. Hamilton, C. Jin, A. van Schaik, and J. Tapson, “An active 2-D
silicon cochlea,” IEEE Trans. Biomed. Circuits Syst., vol. 2, no. 1, pp.
30–43, Mar. 2008.
[14] B. C. J. Moore, An Introduction to the Psychology of Hearing. New
York: Elsevier Academic Press, 2004.
[15] W. S. Rhode and P. H. Smith, “Characteristics of tone-pip response patterns in relationship to spontaneous rate in cat auditory-nerve fibers,”
Hearing Res., vol. 18, no. 2, pp. 159–168, 1985.
[16] B. Wen and K. A. Boahen, “Active bidirectional coupling in a cochlear
chip,” in Advances in Neural Information Processing Systems, Y.
Weiss, B. Schölkopf, and J. Platt, Eds. Cambridge, MA: MIT Press,
2006, vol. 18, pp. 1497–1504.
[17] B. Wen, “Modeling the nonlinear active cochlea: Mathematics and
analog VLSI,” Ph.D. dissertation, Univ. Pennsylvania, Philadelphia,
2006.
[18] C. D. Summerfield and R. F. Lyon, “ASIC implementation of the lyon
cochlea model,” in Proc. IEEE Int. Conf. Acoust. Speech and Signal
Process., San Francisco, CA, 1992, pp. 673–676.
[19] M. P. Leong, C. T. Jin, and P. H. W. Leong, “An FPGA-based electronic cochlea,” EURASIP J. Appl. Signal Process., vol. 7, pp. 629–638,
2003.
[20] C. A. Mead, Analog VLSI and Neural Systems. Reading, MA: Addison-Wesley, 1989.
[21] Y. Raphael, M. Lenoir, R. Wroblewski, and R. Pujol, “The sensory epithelium and its innervation in the mole rat cochlea,” J. Comp. Neurol.,
vol. 314, pp. 367–382, 1991.
[22] I. Russell and K. Nilsen, “The location of the cochlear amplifier: Spatial
representation of a single tone on the guinea pig basilar membrane,” in
Proc. Nat. Acad. Sci., 1997, vol. 94, pp. 2660–2664.
[23] C. D. Geisler, From Sound to Synapse: Physiology of the Mammalian
Ear. Oxford, U.K.: Oxford University Press, 1998.
[24] C. R. Steele, G. Baker, J. Tolomeo, and D. Zetes, “Electro-mechanical
models of the outer hair cell,” in Proc. Int. Symp. Biophysics of Hair
Cell Sensory Systems, H. Duifhuis, J. W. Horst, P. van Dijk, and S. M.
van Netten, Eds., Singapore, 1993, pp. 207–214.
[25] C. D. Geisler and C. Sang, “A cochlear model using feed-forward
outer-hair-cell forces,” Hear. Res., vol. 86, pp. 132–146, 1995.
[26] K. A. Boahen and A. G. Andreou, “A contrast sensitive silicon retina
with reciprocal synapses,” in Advances in Neural Information Processing Systems, J. E. Moody and R. P. Lippmann, Eds. San Mateo,
CA: Morgan Kaufmann, 1992, vol. 4, pp. 764–772.
[27] A. G. Andreou and K. A. Boahen, “Translinear circuits in subthreshold
MOS,” J. Analog Integr. Circuits Signal Process., vol. 9, no. 2, pp.
141–166, 1996.
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
WEN AND BOAHEN: SILICON COCHLEA WITH ACTIVE COUPLING
[28] D. R. Frey, “Log-domain filtering: An approach to current-mode filtering,” Proc. Inst. Elect. Eng., G Circuits Devices Syst., vol. 140, no.
6, pp. 406–416, 1993.
[29] K. Zaghloul and K. A. Boahen, “An on-off log-domain circuit that
recreates adaptive filtering in the retina,” IEEE Trans. Circuits Syst.
I, Reg. Papers, vol. 52, no. 1, pp. 99–107, Jan. 2005.
[30] T. Y. W. Choi, B. E. Shi, and K. A. Boahen, “An on-off orientation
selective address event representation image transceiver chip,” IEEE
Trans. Circuits Syst. I, Reg. Papers, vol. 51, no. 2, pp. 342–353, Feb.
2004.
[31] W. K. Chen, VLSI Handbook. Boca Raton, FL: CRC, 2000.
[32] T. Delbrück, “Bump circuits for computing similarity and dissimilarity
of analog voltages,” in Proc. Int. Joint Conf. Neural Networks, Seattle,
WA, 1991, vol. 1, pp. I–475–479.
[33] M. C. Liberman, L. W. Dodds, and S. Pierce, “Afferent and efferent innervation of the cat cochlea: Quantitative analysis with light and electron microscopy,” J. Comp. Neurol., vol. 301, no. 3, pp. 443–460, 1990.
[34] C. L. Adamson, M. A. Reid, Z. L. Mo, J. Browne-English, and R. L.
Davis, “Firing features and potassium channel content of murine spiral
ganglion neurons vary with cochlear location,” J. Comp. Neurol., vol.
447, no. 4, pp. 331–350, 2002.
[35] J. H. Wittig and K. A. Boahen, “Silicon neurons that phase-lock,” in
Proc. IEEE Int. Symp. Circuits and Systems, 2006, pp. 4535–4538.
[36] J. Santos-Sacchi, “Voltage-dependent ionic conductances of type-I
spiral ganglion-cells from the guinea-pig inner-ear,” J. Neurosci., vol.
13, no. 8, pp. 3599–3611, 1993.
[37] E. J. Moore, D. B. Hall, and T. Narahashi, “Sodium and potassium currents of type I spiral ganglion cells from rat,” Acta Oto-Laryngologica,
vol. 116, no. 4, pp. 552–560, 1996.
[38] K. Yamaguchi and H. Ohmori, “Voltage-gated and chemically gated
ionic channels in the cultured cochlear ganglion neurone of the chick,”
J. Physiol., vol. 420, pp. 185–206, 1990.
[39] J. Lazzaro, M. Wawrzynek, M. Mahowald, M. Sivilotti, and D. Gillespie, “Silicon auditory processors as computer peripherals,” IEEE
Trans. Neural Netw.,, vol. 4, no. 3, pp. 523–528, May 1993.
[40] K. A. Boahen, “A burst-mode word-serial address-event channel-I:
Transmitter design,” IEEE Trans. Circuits Syst. I, Reg. Papers, vol. 51,
no. 7, pp. 1269–1280, Jul. 2004.
[41] V. Chan, S.-C. Liu, and A. van Schaik, “AER EAR: A matched silicon
cochlea pair with address event representation interface,” IEEE Trans.
Circuits Syst. I, vol. 54, no. 1, pp. 48–59, Jan. 2007.
[42] C. A. Mead and T. Delbrück, “Scanners for visualizing activity of
analog VLSI circuitry,” J. Analog Integr. Circuits Signal Process., vol.
1, pp. 93–106, 1991.
[43] M. A. Ruggero, N. C. Rich, S. S. Narayan, and L. Robles, “Basilar
membrane responses to tones at the base of the chinchilla cochlea,” J.
Acoust. Soc. Am., vol. 101, no. 4, pp. 2151–2163, 1997.
[44] L. Robles and M. A. Ruggero, “Mechanics of the mammalian cochlea,”
Physiol. Rev., vol. 81, no. 3, pp. 1305–1352, 2001.
[45] A. van Schaik, E. Fragniére, and E. A. Vittoz, “Improved silicon
cochlea using compatible lateral bipolar transistors,” in Advances in
Neural Information Processing Systems, D. S. Touretky, M. C. Mozer,
and M. E. Hasselmo, Eds. Cambridge, MA: MIT Press, 1996, vol.
8, pp. 671–677.
[46] A. van Schaik and E. Fragniére, “Pseudo-voltage-domain implementation of a design of two-dimensional silicon cochlea,” in Proc. IEEE
Int. Symp. Circuits and Systems, 2001, vol. 3, pp. 185–188.
455
[47] P. R. Kinget, “Device mismatch and tradeoffs in the design of analog
circuits,” IEEE J. Solid-State Circuits, vol. 40, no. 6, pp. 1212–1224,
Jun. 2005.
[48] B. Linares-Barranco, T. Serrano-Gotarredona, R. Serrano-Gotarredona, and G. Vicente-Sanchez, “On mismatch properties of MOS
and resistors calibrated ladder structures,” in Proc. IEEE Int. Symp.
Circuits and Systems, 2004, vol. 1, pp. 377–380.
[49] S. T. Neely, “Finite-difference solution of a two-dimensional mathematical model of the cochlea,” J. Acoust. Soc. Am., vol. 69, no. 5, pp.
1386–1393, 1981.
Bo Wen (SM’05–M’06) received the B.S. and M.S.
degrees in electrical engineering from the University
of Science and Technology Beijing, China, in 1995
and 1999, respectively, and the Ph.D. degree in neuromorphic engineering from the University of Pennsylvania, Philadelphia, in 2006.
Her research interest is investigating neural
coding/mechanisms of sound and speech signals
in the auditory system through computational
modeling—including computer software and
mixed-signal VLSI circuits—and neurophysiological recordings.
Kwabena Boahen (M’06) received the B.S. and
M.S.E. degrees in electrical and computer engineering from the Johns Hopkins University,
Baltimore, MD, in 1989, and the Ph.D. degree in
computation and neural systems from the California
Institute of Technology, Pasadena, in 1997.
Currently, he is an Associate Professor in the Bioengineering Department at Stanford University, Stanford, CA. He is a Bioengineer who uses silicon integrated circuits (ICs) to emulate the way neurons compute, linking the seemingly disparate fields of electronics and computer science with neurobiology and medicine. His contributions to the field of neuromorphic engineering include a silicon retina that could
be used to give the blind sight and a self-organizing chip that emulates the way
the developing brain wires itself up. His scholarship is widely recognized, with
many publications to his name, including a cover story in the May 2005 issue
of Scientific American.
Dr. Boahen has received several distinguished honors, including a Fellowship from the Packard Foundation in 1999, a CAREER award from the National Science Foundation in 2001, a Young Investigator Award from the Office
of Naval Research in 2002, and the National Institutes of Health Director’s Pioneer Award in 2006. From 1997 to 2005, he was on the faculty of the University
of Pennsylvania, Philadelphia, where he held the first Skirkanich Term Junior
Chair.
Authorized licensed use limited to: Stanford University. Downloaded on December 28, 2009 at 17:53 from IEEE Xplore. Restrictions apply.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement