Sensorwebs: 2001 DARPA Summary

1
Distributed sensor networks opportunities and
challenges in signal processing and communications
Kannan Ramchandran EECS Department BASiCS
Research Group U.C. Berkeley
Berkeley Audiovisual Signal processing
Communication Systems
2
Talk Outline

• Overview of related research activities
• Vision on future research directions

3
Sensorwebs _at_ UC Berkeley

• Interdisciplinary project Creation of a
  fundamental unifying framework for real-time
  distributed information processing with
  applications to sensornets, consisting of
• Distributed signal processing
• Distributed control
• MEMS
• Multiterminal information theory
• Distributed learning theory

4
Distributed SP/Comm overview
(DISCUS)

Distributed compression to seamlessly exploit
relevant correlations

Robust estimation in rate
-
constrained
unreliable
sensor networks
Distributed sampling theory for dense sensor
networks


Duality
between
distributed compression
and
data
-
embedding/security

High Capacity multimedia data
-
hiding/spectrum recycling/
steganography
http//www.basics.eecs.berkeley.edu/
jimchou
/
researchlinks
/
audiohiding
/audio.html

VISDOM

V
ideo
S
treaming using
D
istributed encoding
O
ver
M
ultiple servers
http//www.basics.eecs.berkeley.edu/rpuri/researc
hlinks/
visdom
/
visdom
.html

PRISM

P
ower
-
efficient, h
I
gh
-
compression,
S
yndrome
-
based
M
ultimedia coding
http//www.basics.eecs.berkeley.edu/rpuri/researc
hlinks/
prism/prism.html
5
Distributed compression for sensor networks

• Nodes X, Y have correlated data.
• X-Y communication is expensive.
• Can we exploit correlation without communicating?
• Information-theory gives asymptotic answers
• Slepian-Wolf, Wyner-Ziv theorems
• Constructive framework based on coding and
  quantization theory DISCUS (DIstr. Source Coding
  Using Syndromes)
• Nested quantization/modulation codes based on
  trellises, LDPCs, turbo codes ? 1-2 dB from
  W-Z bound
• Possible to integrate correlation tracking and
  distributed coding in some cases, e.g. for audio
  applications

Y
X
Dense, low-power sensor-networks
6
Distributed coding for audio rendering
Yi hiX Zi Zi Noise hi LTI
Filter
7
Robust estimation under rate-constraints Gau
ssian sensor network

• n sensors transmit noisy observations of X over
  rate-constrained (R) channels
• Channel delivers some arbitrary k sensor
  packets.
• Question what is the best achievable MMSE
  estimation quality?

8
Robust estimation under rate constraints

• Optimal estimation performance for (k,k)
  reliable sensor network case
• (i.e. n k case known Oohama 98)
• For unreliable case kltn (i.e., an (n,k) sensor
  network) (Puri, Ramchandran et al. 02)
• can match the optimal performance of the
  reliable (k,k) sensor network!
• robustness comes without loss of estimation
  performance!
• Key idea Quantize observations Yi using
  distributed compression principles

9
Main Result

• For a (k, k) sensor network, complete
  rate-distortion region known (Oohama 1998).

Quality
(Rate)
kR
10
Main Result

• For a (k, k) sensor network, complete
  rate-distortion region known (Oohama 1998).
• For an (n, k) sensor network, can match above
  performance for the reception for
• any k packets and delivers better quality for
  reception of more packets!
• Robustness for no loss in performance.

11
Distributed sampling for dense sensor networks

• Conservation of bits principle Tradeoff
  between oversampling rate and A/D converter
  precision (A. Kumar, Ishwar Ramchandran,
  submitted IPSN 03)
• Local communication model

12
Main idea 1 bit sensor A/D example

• Nyquist rate at R bit A/D precision ?error
• R-times oversampled at 1 bit A/D ? error
• Distributed processing Nearest-neighbor
  single-bit comm. single sensor comm. to central
  unit
• Dither signal d(t) should be smooth and force
  zero-crossing per Nyquist interval
• Dither d(t) can be kept secret to provide security

13
Some relevant higher-level directions

• Fundamental bounds in large-scale sensor
  networks
• scaling laws (Gupta and Kumar)
• extensions to correlated sensor data models,
  mobility, etc.
• phase-transition phenomena (statistical physics)
• multi-terminal information theoretic bounds for
• Network channel coding multi-access, broadcast,
  relays
• Network source coding multiple-description,
  distr. source coding
• NSCC end-to-end metric under system power
  constraints
• Cross-network layer optimization
• no legacy requirements higher impact potential
• addressing/ packetization/synchronization/
  aggregation/
• Incorporating security/privacy holistically
• Closing the loop around the data distributed
  control theory

14
Distributed DSP (DSP) simplistic view

• Revisit many classical SP problems (estimation,
  inference, detection, classification, fusion)
  under constraints of
• bandwidth (compression)
• noisy transmission medium (channel coding)
• total system energy (communication processing)
• highly unreliable system components (fault
  tolerance)

15
Communication vs. processing power tradeoffs
(\citeKris_Pister)

• Mica (macro smart dust) http//www.cs.berkeley
  .edu/polastre/papers/wsna02.pdf
• 4 MHz CPU taking 5 mA current _at_ 3V ? 3.75
  nJ/cycle
• Tx. 20 nAh/packet _at_ 3V 60 nW3600s/pkt.220mi
  croJ/pkt.
• 7 microJ/bit
• Rx. 8 nAh/packet ?3 microJ/bit
• Smart dust projections (cubic mm size scale)
• 10 pJ/inst
• Tx.Rx. ?1 nJ/bit
• Rockwell WINS nodes
• 1500-2700 instructions per Tx. bit
• Medusa II nodes (UCLA)
• 220-2900 instructions per Tx. bit

16
Distributed DSP (DSP) contd.

• distributed and robust estimation, inference,
  detection, classification, fusion, sampling
  theory
• integration with distributed learning theory
• graphical models exploit local structure in
  global problems
• incorporate learning seamlessly into the loop?
• independent and dependent component analysis
• fast robust algorithms for signal separation
• incorporating security
• fundamental duality between distributed
  compression and covert communications (coding
  with side information)
• interdisciplinary approach to cryptography/SP/com
  m.
• joint encryption and compression
• secure sampling
• use of error-correction-codes for robust coding
  and distributed secret sharing, etc.
• identifying isomorphisms with related problems
  in
• image processing/computer vision/wavelet
  theory/coding theory/

17
Analogies and isomorphisms

• parallel computing
• sensors ??processors
• individual node capabilities fault-tolerance
  issues may be exaggerated for sensor network
  problem
• computer vision
• tracking ?? segmentation
• image and signal processing
• network ??image, sensors ?? pixels
• coding theory
• codes on graphs (Yeung et al., Medard Kotter)
• Wavelet theory
• multiscale processing, localized processing
• Distributed array processing
• antenna array elements ??sensor nodes

18
Distributed DSP (contd.)

• Need fundamental paradigm shift in design
  architectures
• Asymmetric complexities
• shift burden from remote units to centralized
  node
• Robustness fault-tolerant designs
• Diversity in representation/communication
• Rehaul prediction-based frameworks LP, DPCM,

19
Example Rethinking uplink video coding
(Puri Ramchandran, 02
University of California, Berkeley
20
Physical layer challenges

• low-power unreliable radios
• Heterogeneous mix of many cheap devices a few
  reliable ones?
• exploit spatial density of sensor network for
  channel diversity
• multi-user diversity
• MIMO wireless channels under local communication
  cost constraints
• asynchronous versus synchronous communication
  modes
• hybrid analog/digital communications
• integrated channel estimation/synchronization/pac
  ketization/routing functionalities
• What physical layer technologies? UWB? Analog
  mode?
• Opportunity to influence design architectures
• asymmetric transmitter/receiver complexity
• multi-terminal diversity

21
Questions and caveats

• What are we trying to do with the sensor
  networks?
• Raw data versus low-dimensional feature or
  parameter
• Overly generic solutions vs. being too
  application-oriented?
• Do we really have a handle on the global cost
  function?
• What is network life? Maximum time for
  network to fail?
• How do we quantify network system performance?
• Need to invest as much creative energy in design
  and architecture as for analysis.
• We tend to be too analysis-oriented as a
  community
• Rich supply of generic problems that are
  well-studied in our community
• Inference, detection, estimation, compression,
  classification
• Can be cleanly posed under constraints of
  distributed communication and system energy.
• Oodles of fundamental questions to be
  answered NSF perfect!
• Opportunity to shape the field to some extent
  in the way we want