The BASiCS Group Berkeley Audio-visual Signal processing and Communication Systems

1. The BASiCS Group Berkeley Audio-visual Signal processing and Communication Systems Distributed signal processing: compression: challenges and opportunities Kannan Ramchandran http://www.basics.eecs.berkeley.edu UCB Sensor Nets Day

2. Towards a System Theory for Robust Large-Scale Sensor Networks Design guidelines for robust large-scale networks: Channel Physics, Percolation Theory Representation and Data-Acquisition: Distributed Sampling Theory f(x) Information Dissemination: Routing, Compressing, Mobility Closing the Loop: Inference & Adaptive Control NSF Sensors (Ramchandran, Sastry, Tse, Vetterli, Poolla) UCB Sensor Nets Day

3. Sensor networks: a systems view Systems tasks: • Data acquisition • Distributed compression and communication • Networking and routing • Distributed inference and decision (classification / estimation) • Closing the loop (control) Guiding principles: • Statistical models for sensor-fields • Scaling laws for dense networks • Information and coding theory • Learning theory and adaptive signal processing UCB Sensor Nets Day

4. Distributed SP (DSP): “low-hanging fruit” • Revisit many classical SP problems (estimation, inference, detection, fusion) under constraints of: • bandwidth (compression) • noisy transmission medium (coding + MAC) • total system energy (communication + processing) • highly unreliablesystem components (robust design) • Voila  you get a “distributed signal processing” recipe! • Constraints force robust distributed solutions – sampling, processing, routing, compressing, coding, controlling. • Architectures should reflect and exploit computational diversity in wireless devices (TV’s, cell phones, laptops, cheap sensors) • Asymmetric complexities • In-built robustness & fault-tolerant designs: • Diversity in representation & communication • Rehaul “deterministic” frameworks (e.g. prediction-based) with “probabilistic” ones UCB Sensor Nets Day

5. bad good good “Central unit” Sampling sensor fields • Many physical signals e.g., pressure, temperature, are approximately BL • Physical propagation laws often provide a natural smoothing effect A/D converters (sensors) • Sensor network constraints • Low-precision A/D • Limited power and bandwidth Sampling a 1-D spatio-temporal field space 2X 2X 2X X X X T 3T 2T time UCB Sensor Nets Day

6. f(x) Motivation: Acquisition & reconstruction of sensor fields • Is there an “information” scaling law ? • [Gupta-Kumar’00]: In ad-hoc networks, with independent data sources, throughput/sensor  0 as 1/sqrt(N). • In sensor nets, data correlation increases with density. • Can information-rate/sensor and reconstruction distortion • go to zero with density? • Tradeoffs between sensor precision and # of sensors? • Can we overcome low precision sensors by throwing • scale at the problem? • Is there an underlying “conservation of bits” principle? UCB Sensor Nets Day

7. Sensor-Field Reconstruction: ‘Distributed’ Sampling Theory Error  similar D = c 2-k accuracy D k Bit-budget  (1,k) (2,k-1) (k-2,3) log (# of sensors) (k-1,2) 1 bit/sample, T/2 2 bits/sample, T (k,0) A/D precision b-bit • Distortion ~ O(1/N) • RNyquist~ O(log N) • Rsensor ~ O(log N / N) • Need concept of “dithering” and “distributed coding” • Ishwar, Kumar & Ramchandran (IPSN ’03) • “Conservation of bits” principle We can trade off A/D precision for oversampling rate (quality bits per Nyquist interval). 0 UCB Sensor Nets Day

8. Can we overcome cheap radios by throwing scale at the problem? • Can we devise clever probabilistic distributed algorithms for routing & network coding that exploit the randomness in the manufacturing process? P(fcarrier) Signal BW 3s variation (Picoradio) fcarrier Overcoming Unreliable Radios • Narrowband Radios • Simple, used by all sensor nodes today [Motes, PicoRadio, Ember, SmartDust] • How to get fcarrier? • Crystal Oscillator (precise but expensive) • MEMS Resonator (less precise & less expensive) • On-chip LC-Resonator (cheap, low-power, imprecise) UCB Sensor Nets Day

9. ^ Y X Encoder Decoder X X Y Information theory:X can be theoretically compressed at a rate equal to that when the encoder too has access to Y Distributed compression Dense, low-power sensor-networks • The encoder needs to compress the source X. • The decoder has access to correlated side • information Y. • Can we compress X to H(X|Y)? Can design practical distr. source coding framework to approach this. UCB Sensor Nets Day

10. R R R X Source Channel Collector Integrating learning: correlation tracking • Many sensors report to controller • Correlation tracking • Controller keeps track of correlation • Specifies how much compression • Sensors blindly encode readings • Minimal processing at sensor nodes • Complexity at controller • Cheap sensors • Probabilistic reference to side information allows for robustness to packet loss UCB Sensor Nets Day

11. Collaborative processing: compressing raw-data versus local estimates • Several scenarios: • Sensor-clusters (groups of sensors that can collaborative) • Multiple antennas per sensor • Multimodal sensors UCB Sensor Nets Day

12. Result • If collaborative processing is (MSE) optimal when R is infinity, … • Here, R = infinity and UCB Sensor Nets Day

13. Result • … then it is also optimal for any finiteR. Suggests that distributed estimation and compression tasks can be “de-coupled”, i.e., one can design & adapt network topology by ignoring bandwidth requirements in a number of scenarios. UCB Sensor Nets Day

14. Opportunities: architecture rehauls • Architectures should reflect and exploit computational diversity in wireless devices (TV’s, cell phones, laptops, cheap sensors) • Asymmetric complexities • In-built robustness & fault-tolerant designs: • Diversity in representation & communication • Rehaul “deterministic” frameworks (e.g.prediction-based frameworks: LP, DPCM, etc.) with “probabilistic” ones UCB Sensor Nets Day

15. Wireless Network • Changing landscape: “uplink” heavy applications • Ultra-low-power video sensors and surveillance cameras • Multimedia-enabled cellphones & PDA’s • High-resolution wireless digital video cameras • Wireless-video teleconferencing systems • Home-entertainment and home-networking systems Video is not just a downlink broadcast experience any more! Rethinking video-over-wireless • Today’s video architectures shaped by downlink broadcast model: • Complex encoder • Light decoder Motion estimation task dominates (up to 90%) UCB Sensor Nets Day

16. New class of video codecs: requirements • Light codec complexity in order to • Maximize battery-life. • Satisfy complexity constraints at encoding device. • High compression efficiency to match • Available bandwidth/storage constraints. • Low transmission power constraints. • Robustness to packet/frame drops to • Combat harsh wireless transmission medium. UCB Sensor Nets Day

17. Decoder Encoder Transcoding proxy light Heavy encoder Light decoder Rethinking the division of labor Under reasonable signal models, it ispossible to transfer (motion search) complexity to decoder without loss of compression efficiency (Ishwar, Prabhakaran, & Ramchandran, 2003) UCB Sensor Nets Day

18. PRISM video simulation results • Sequence used: Football (14 frames, 352x240) • Comparison: H.263+ (free version from UBC, Vancouver) • Frame rate: 30fps, Encoding rate: 10kB per frame • Compression: Performance is visually competitive with respect to full-motion complex inter-frame codecs such as MPEG-4 & H.263+. • (For pure compression, H.263+ outperforms PRISM by about 1.3 dB on our tests on the Football sequence) • Robustness: Much more robustthan current solutions. Can recover from frame losses. • Test for robustness: second frame was removed from frame memory after decoding. third frame was decoded off the first frame in both cases. UCB Sensor Nets Day

19. Qualcomm’s simulator for “CDMA-2000 1X” • At packet error rate 6%: • At packet error rate 11%: • H.263+ at packet error rate of 3% and PRISM at 16%: PRISM is 4-8 dB better than H.263+ for the loss rates investigated. UCB Sensor Nets Day