Académique Documents
Professionnel Documents
Culture Documents
Brain-inspired auditory •
Can spike-based speech
recognition systems outperform
conventional approaches?
Recently, several approaches for learning or neuro-muscular systems. Our initial simula- #6
Output
in SNNs have been proposed.3 Here, we tions in this area point out the suitability of #5
#4
will focus on one called ReSuMe2,6 (remote ReSuMe as a training method for SNN-based
#3
supervised method) that corresponds to the neurocontrollers in movement generation and #2
Widrow-Hoff rule and is well known from control tasks.8,9,11 #1
traditional artificial neural networks. ReSuMe Real-life applications of SNN require effi- 0 0.05 0.1 0.15 0.2 0.25 0.3
takes advantage of spike-based plasticity cient hardware implementations of the spiking results after 5 training epochs
mechanisms similar to spike-timing depen- models and the learning methods. Recently, # 10
dent plasticity (STDP).1,6 Its learning rule is ReSuMe was tested on an FPGA platform:4 #9
Output
#5
where Sd(t), Sin(t) and So(t) are the desired pre- processing ability, the system is able to meet
#4
and postsynaptic spike trains,1 respectively. the time restrictions of many real-time tasks.
#3
The constant a represents the so-called non- #2
Hebbian contribution to the weight changes. Filip Ponulak #1
The role of this parameter is to adjust the Inst. of Control and Information Eng. 0 0.05 0.1 0.15 0.2 0.25 0.3
average strength of the synaptic inputs so as Posnan University of Technology results after 10 training epochs
to impose on a neuron the desired level of Posnan, Poland # 10
activity (desired mean firing rate). The func- E-mail: Filip.Ponulak@put.poznan.pl #9
Hynek Hermansky
IDIAP Research Institute
Swiss Federal Institute of Technology
Lausanne, Switzerland
Email: hynek.hermansky@idiap.ch
Figure 3. Illustration of the technique for obtaining a reliable estimate of posterior probability
density functions pi(Q|X) without the use of top-down constraints L. The short-term critical-band References
1. H. Hermansky and N. Morgan, Automatic Speech
spectrogram (left part of the figure) is derived by weighted summation of appropriate components Recognition, in Encyclopedia of Cognitive Science,
of the short-term spectrum of speech. A segment of this spectrogram is projected on 448 different L. Nadel, Ed., Nature Publishing Group, Macmilian
time-frequency bases (shown in Figure 3), centred at the time instant i, yielding a 448 point Publishers, 2002.
2. Bourlard, H. and Morgan, N., Connectionist Speech
vector that forms the input to the MLP neural net, trained on about 2 hours of hand-labelled
Recognition—A Hybrid Approach, Kluwer Academic
telephone-quality speech to estimate a vector of posterior probabilities pi(Q|X). A set of pi(Q|X) Publishers, 1994.
for all time instants forms the so-called posteriogram, shown for the utterance one-one-three-five- 3. H. Bourlard, C. J. Wellekens, Links between Markov
eight in the lower part of the figure. Higher posterior probabilities are indicated by warmer colors Models and Multilayer Perceptrons, IEEE Conf. Neural
Information Processing Systems, 1988, Denver, CO,
(see Reference 5 for more details). Ed. D. Touretzky, Morgan-Kaufmann Publishers, pp.
502-510, 1989.
4. H. Ketabdar and H. Hermansky, Identifying and
dealing with unexpected words using in-context and out-of-
context posterior phoneme probabilities, IDIAP Research
Report, 2006.
5. H. Hermansky and P. Fousek, Multi-resolution RASTA
filtering for TANDEM-based ASR, in Proc. Interspeech
2005, 2005.
Hermansky, continued p. 10
Embedded vision, continued from p. This algorithm, implemented in the Pierre-François Rüedi and Eric Grenet
BlackFin processor, works robustly at 25 CSEM S.A.
sisting of two edges with high contrast frames per second in varying conditions Neuchâtel, Switzerland
magnitude and opposite contrast direc- such as night, sun in the field of view, and E-mail: pfr@csem.ch, egt@csem.ch
tions—is detected and tracked in a restricted roads with poor quality markings. For
area that is continuously adapted to the last demonstration purposes, detection results References
detected position. Continuous and dashed (mark position and type, road curvature, 1. M. Barbaro, P.-Y. Burgi, A. Mortara, P. Nussbaum
markings are differentiated. The vanishing light level, etc.) are sent via the low-data-rate and F. Heitger, A 100×100 pixel silicon retina for gradi-
ent extraction with steering filter capabilities and temporal
point is extracted, the variations of which radio-frequency link to a cellular phone that
output coding, IEEE J. Solid-State Circuits 37 (2),
give useful gyroscopic information (tilt and displays a synthetic view of the road in real pp. 160-172, 2002.
yaw angles). A Kalman filter supervises the time (see Figure 3). 2. P.-F. Rüedi, P. Heim, F. Kaess, E. Grenet, F. Heitger,
system and gives robustness to the detec- This work demonstrates that moving P.-Y. Burgi, S. Gyger, P. Nussbaum, A 128×128 Pixels
120 dB Dynamic Range Vision Sensor Chip for Image Con-
tion (e.g. when markings are temporarily some of the image processing to the sensor trast and Orientation Extraction, IEEE J. Solid-State
missing). The system also estimates the itself is a solution to implement real-time Circuits 38 (12), pp. 2306-2317, Dec. 2003.
illumination level and road curvature by low-power and low-cost vision systems 3. D. J. Field, What is the goal of sensory coding?, Neural
Computation 6, pp. 559–601, 1994.
fitting the markings points with a clothoid able to function robustly in uncontrolled
4. E. Grenet, Embedded High Dynamic Vision System For
equation, allowing it to appropriately con- environments. Real-Time Driving Assistance, TRANSFAC ’06, San
trol the headlights. Sebastian, Spain, p. 120, October 2006.
Figure3. Various road situations and their related symbolic representation. Shown are single-lane (top left) and a
multi-lane curves (top middle) by day, a lane departure in a tunnel (bottom middle) and on a countryside road
with single marking by night (bottom left). To the right is a real-time display and a warning on a cell phone.
5
10
First formant frequency at 426 Hz
0
10
−5
10
0 1000 2000 3000 4000 5000 6000
Frequency in Hz
FFT Magnitude of ISI Histogram for the Channel at 437 Hz
600
400 Second peak at 428 Hz
DoS
0 200 400 600 800 1000 1200 1400 1600 1800 2000
Frequency in Hz
Figure 1. Log-magnitude spectral envelope for /uh/ and the corresponding degree of phase synchrony for two sets of
hair cells centered at 437Hz and 519Hz (computed for a noisy utterance with 5dB SNR).
DoS at 248Hz
Figure 1, even with a very noisy vowel input
signal, the fibers with characteristic frequencies
(437Hz) close to the first peak (426Hz) in the
vowel’s log-magnitude spectral plot are still
Neural Readout able to phase lock very close to that particular
Circuit Function
frequency. They also have a higher DoS than
Logarithmically
other channels, such as the one shown in the
Distributed
bottom plot with a characteristic frequency
Channels
(519Hz) further from the first formant peak.
Finally, for classification, the system
employs an LSM with a randomly connected
DoS at 4kHz
recurrent neural circuit.3 The idea is to map
the input vector to a higher dimension where
the distance metric between prospective
classes is larger. For our system, the input
: Randomly distributed dynamic spiking synapses vector—which is comprised of the degrees of
synchrony for each channel—is passed on to
: Low pass filtering and sampling
the neural circuit as the membrane potentials
of input neurons that make dynamic spiking
Figure 2. The overall spike-based classification. The degree of synchrony is extracted from spike synapses with the circuit using spike-timing
trains generated in each individual cochlear channel. This feature set is then used with an LSM dependent plasticity. The state of the circuit is
with supervised learning for classification. low-pass filtered and sampled to be associated
Table 1. with a target class (different types of vowels) by
the help of a trainable readout function. Figure
SNR (dB) 25 10 5 2 shows the overall system design, as well as
some of the important system parameters.
Also, see one of the highlights of the 2005 workshop, “The Grand Challenge” at:
http://www.youtube.com/watch?v=Gs9P35Fq3Gw
Spike-based speech, continued from p. tonotopic neuron populations even under the References:
presence of high amounts of noise. 1. I. Uysal, H. Sathyendra, and J. G. Harris, A biologically
plausible system approach for noise robust vowel recognition,
results are shown in Table 1. Future work involves extrapolation of IEEE Proc. of MWSCAS, CD-ROM, 2006.
At high signal-to-noise ratio (SNR) these findings to more complex signals and 2. C. J. Sumner, E. A. Lopez-Poveda, L. P. O’Mard, and
values, both systems perform comparably multi-syllable words by the help of relational R. Meddis, Adaptation in a revised inner-hair cell model, J.
Acoust. Soc. Am. 113 (2), p. 893-901, 2003.
well, but the proposed system using phase networks as observed in the cortex.
3. W. Maass, T. Natschlager, and H. Markram, Real-time
synchrony coding is able to outperform the computing without stable states: A new framework for neural
MFCC-HMM algorithm by 12% at 5dB computation based on perturbations, Neural Computation
SNR. In regards to the question raised in the Ismail Uysal, Harsha Sathyendra, and 14 (11), pp. 2531-2560, 2002.
title, though applied to a simplified domain, John G. Harris
spike-based recognition is clearly more noise Computational NeuroEngineering Lab
robust when compared to a conventional ASR University of Florida
system. This performance is mainly due to the Gainesville, FL, USA
phase synchrony maintaining capabilities of E-mail: ismail@cnel.ufl.edu
Ever wondered why progress seems slow in possible solutions, and finally concludes in
building visually guided autonomous agents proposing a general computational architec-
that perceive and intelligently interact with ture for visual motion perception. The key
their environment? Well, one reason may concept is that the perceptual process is an
be that our understanding of perception optimization problem of finding the visual
and the underlining computations involved motion estimate that is maximally consis-
is incomplete or just plain wrong. This tent with the visual information and the
new book by Alan Stocker provides n system’s expectations. Chapter three makes
unconventional and fresh perspectives on the connection to associative memory and
how to understand perception and build Hopfield networks as examples of network
simple artificial perceptual systems using architectures that compute optimal solu-
analog VLSI (very large silicon integra- tions. It demonstrates how simple problems
tion) circuits. Focusing on the example of (e.g. the winner-take-all operation) can be
visual motion perception, it demonstrates formulated as local constraints that together
how brain-style computation combined define the optimal solution. The chapter also
with CMOS (complimentary metal-oxide shows how to derive appropriate network tell if this is true or not.
semiconductor technology can lead to ef- architectures that find it. The broad approach of this book
ficient and robust ‘neuromorphic’ circuits Chapter four then formulates optical- certainly reflects the background and the
to solve the hard optimization problems flow estimation as a constraint satisfac- interests of the author. He is an expert
encountered in perception. tion problem, deriving the basic network aVLSI circuit designer, a computational
One key factor underlying the success architecture that is the basis for all further modeler of the visual system, and a psy-
of human visual perception lies in its use of networks discussed in the book. It draws chophysical experimentalist working on
constraint satisfaction. That is, the brain pre- the connection between the formulated human motion perception. I highly rec-
sumably applies mechanisms that combine constraint solving problem and statistically ommend this book not only to those who
the aspects of its visual input that cohere optimal motion estimation as described are particularly interested in aVLSI visual
and segments out those aspects that do not. with Bayesian frameworks, showing that motion circuits, but to anyone interested in
These mechanisms bootstrap globally coher- prior information is essential in achieving the novel, neuromorphic, style of compu-
ent (optimal) solutions by rapidly satisfying a robust design. Furthermore, extensions tation. The philosophy and methodology
local consistency constraints. Consistency of the basic network allow even more of the approach seem general enough and
depends on relative computations such as sophisticated processing such as motion applicable to other perceptual tasks, such
non-linear comparison, interpolation and segmentation or motion selection for which as depth perception and texture segmenta-
error feedback, rather than absolute preci- the network selects regions in its visual field tion. Furthermore, the analog VLSI imple-
sion. And this style of computation is very that match a particular motion and size. mentation of the presented computational
suitable for implementation in analog VLSI Chapters five to seven extensively deal networks becomes particularly attractive in
circuits, as Dr. Stocker demonstrates. with aVLSI implementations of the pro- light of recent technological developments
What makes this book special is that it posed network architectures, providing de- in three- dimensional integrated circuits.
not only presents practical implementations tailed schematics and measurements of the Three-dimensional integration permits lo-
of constraint satisfaction networks for visual fabricated chips. The effects of the inevitable cal vertical connections between different
motion perception, but it also demonstrates non-linearities and mismatch are discussed chips, physically stacked as a ‘layer cake’.
a series of useful and impressive aVLSI in detail, showing that clever analog designs Recurrent analog networks can naturally
circuits for solving visual motion problems can take advantage of nonlinearities to im- be implemented as multi-layered parallel
such as estimating 2D optical flow, motion prove robustness and performance. computational blocks of tremendous capa-
segmentation, and motion selection. And The book concludes with an interesting bilities, without the need for sophisticated
these chips are useful for robotic applica- final chapter with a comparison to primate chip-to-chip protocols.
tions. Their true strength lies, however, visual motion perception systems. It also
in their broad and principled theoretical presents data of head-to-head comparison Ralph Etienne-Cummings
foundations. between humans and the aVLSI chips per- Department of Electrical and Computer
The book begins with some ecological forming the same perceptual tasks. The dy- Engineering
considerations about why and how visual namics and steady-state behavior similarities Johns Hopkins University
motion is perceived from changes in the are quite surprising, leading the author to Baltimore, MD, USA
visual input. It then goes on to illustrate the conclude that both systems must optimize Email: retienne@jhu.edu
basic computational challenges, discusses a similar set of constraints. The future will URL: http://etienne.ece.jhu.edu/
Figure 2. Demonstration
system for blind signal
processing and adaptive
noise cancellation. Two
microphones received six
signals: one human speech,
one car noise from the
right speaker, and four
background music signals
from the remaining four
speakers.