Get Over the I nsecurity! Ed Lazowska Depart ment of Comput er Science & Engineering Universit y of Washingt on
Key point s � Don’t get hung up on t rying t o be a “pure science” � The f act t hat much of what we do is usef ul is good , not bad � Sur e, t he physicist s did The Mot her Of All Demos back in 1945, but t hey’re in t he crapper t oday – now, t hey envy us! � We are at t he cent er of everyt hing
� There are incredible opport unit ies f or “peer t o peer” int ellect ual advancement � “J ust say no” t o t hose who want somet hing else f rom you – cor por at e or academic � But r ecognize t hat ever y par t y in a collabor at ion needs t o “pay some dues” � Beware of having a narrow view of what const it ut es comput er science
Science vs. engineering � Science � Describe, explain � Engineering � Design, build, evaluat e � “An engineer can do f or a dime what any f ool can do f or a dollar ” � Much of comput er science is engineering – celebrat e t his!
“Engineering research”: oxymoron? � “Fundament al research” and “applicat ion- mot ivat ed research” are compat ible
Tradit ional view Fundamental Applied research research
Alt ernat ive view Pasteur; much of Edison biomedical and Concern with use engineering research Bohr Concern with fundamentals
Some UW examples in t he bio space � Comput at ional molecular biology � LabScape – embedded syst ems t o inst rument biot ech laborat ories � Neurally-inspired comput ing
Comput at ional Molecular Biology � Collabor at or s: Lee Hood, Maynar d Olson, Phil Gr een � Facult y: Dick Karp, Mart in Tompa, Larry Ruzzo, Rimli Sengupt a � Post docs: Amir Ben-Dor, Benno Schwikowski � Complet ed Ph.D. st udent s: Br endan Mumey (U Mont ana) , J er emy Buhler (WashU) , Ka Yee Yeung (UW Microbiology) , Agat ha Liu (I BM) , Saur abh Sinha (Rockaf eller U) , Mat hieu Blanchet t e (McGill) , Emily Rocke (UW Genome Sciences) � Cor por at e int er act ions: Zymogenet ics, I mmunex, Roset t a, I nst it ut e f or Syst ems Biology
The Port olano Port olano Expedit ion Expedit ion The in I nvisible Comput ing in I nvisible Comput ing Gaet ano Bor r iello Gaet ano Bor r iello Depar t ment of CS&E Depar t ment of CS&E Univer sit y of Washingt on Univer sit y of Washingt on Seat t le SAGE Gr oup Seat t le SAGE Gr oup 14 Sept ember 2000 14 Sept ember 2000 port olano port olano.cs cs.washingt on washingt on.edu edu
P rincipal Themes P rincipal Themes � I nvisibilit y � I nvisibilit y � not enough t o be mobile, pervasive, ubiquit ous, et c. not enough t o be mobile, pervasive, ubiquit ous, et c. � user’s at t ent ion is t he valuable resource user’s at t ent ion is t he valuable resource � minimize user conf igurat ion/ maint enance/ int eract ion minimize user conf igurat ion/ maint enance/ int eract ion � robust , reliable, saf e, and t rust wort hy robust , reliable, saf e, and t rust wort hy ware, and “applicat ions” � � services � devices, middle devices, middle- -ware, and “applicat ions” services � Act ive f abric � Act ive f abric � plug plug- -and and- -play, discovery, play, discovery, composabilit y composabilit y � dat a dat a- -cent r ic, het er ogeneous, act ive net wor king cent r ic, het er ogeneous, act ive net wor king � dat a and code mobilit y dat a and code mobilit y � self self - -or ganizing, self or ganizing, self - -updat ing, self updat ing, self - -monit oring syst ems monit oring syst ems � act ive dat abases and inf ormat ion management act ive dat abases and inf ormat ion management � Ext ernal user communit y � Ext ernal user communit y
LabScape - - one of our driver applicat ions one of our driver applicat ions LabScape � Biology is a hard science wit h a sof t inf rast ruct ure � Biology is a hard science wit h a sof t inf rast ruct ure � capt ure and use of knowledge is key � capt ure and use of knowledge is key � f rom loosely connect ed t o highly int egrat ed collaborat ion � f rom loosely connect ed t o highly int egrat ed collaborat ion � invisible inf rast ruct ure f or building knowledge base � invisible inf rast ruct ure f or building knowledge base Hypot hesize Experiment knowledge base knowledge base I nt erpret Descript ive Experiment publicat ion Model Manager Hypot hesize Experiment I nt erpret
Event Capt ure in Labscape Event Capt ure in Labscape
Neurally Inspired Inspired Neurally Computation Computation Chris Diorio Computer Science & Engineering University of Washington diorio@cs.washington.edu
Nature is telling us something... � Can add numbers together in � Can understand speech trivially � Far ahead of digital computers nanoseconds � …and Moore’s law will end � Hopelessly beyond the capabilities of brains
Problem: How do we build circuits that learn � One approach: Emulate neurobiology � Dense arrays of synapses error signal learn signal synapse synapse W 21 W 22 output = ∑ W X 2 j j j error signal learn signal synapse synapse W 11 W 12 output = ∑ W X 1 j j j X 1 input vector X X 2
Silicon synapses � Use the silicon physics itself for learning � Local, parallel adaptation � Nonvolatile memory Silicon Synapse Transistor Charge Q Sets the Weight 10 -5 10 -6 source current (A) Q 1 Q 2 10 -7 Q 3 10 -8 Q 4 Q 5 10 -9 n + n + n + p n – 10 -10 floating gate electron electron injection tunneling (charge Q ) 10 -11 p – substrate 0 1 2 3 4 5 control-gate–to–source voltage (V)
Silicon synapses can mimic biology � Local, autonomous learning Biological Synapses Silicon Synapses 5 synapse source currents (nA) 4 3 2 1 0 –10 0 10 20 30 40 50 time (min) Mossy-fiber EPSC amplitudes plotted over time, before and after the Synapse transistor source currents plotted over time, before and after we applied a tetanic stimulation of 2×10 5 coincident (row induction of LTP. Brief tetanic stimulation was applied at the time in- dicated. From Barrionuevo et al., J. Neurophysiol. 55:540-550, 1986. & column) pulses, each of 10 µs duration, at the time indicated.
Synaptic circuits can learn complex functions 1 � Synapse-based circuit operates on probability distributions 0.8 true means � Competitive learning circuit output value (V) � Nonvolatile memory 0.6 software neural � 11 transistors network 0.4 � 0.35µm CMOS � Silicon physics learns 0.2 “naturally” 0 1000 2000 3000 4000 number of training examples � Silicon learning circuit versus software neural network � Both unmix a mixture of Gaussians � Silicon circuit consumes nanowatts � Scaleable to many inputs and dimensions
Technology spinoff: Adaptive filters � Synapse transistors for signal processing � ~100× lower power and ~10× smaller size than digital Mixed-signal FIR filter FIR filter with on-chip learning 16-tap, 7-bits 225MHz, 2.5mW 64 taps, 10 bits, 200MHz, 25mW Built and tested in 0.35µm CMOS In fabrication in 0.35µm CMOS Adjust synaptic tap weights off-line On-line synapse-based LMS
Startup company: Impinj � Chris Diorio (UW) and Carver Mead (Caltech) � Self-tuning analog computing implemented in standard digital CMOS processes (e.g., TSMC) for telecommunications applications (filtering, DSP, etc.) � Potentially a factor of 500 power savings, plus the ability to fully integrate analog and digital on the same die
Problem: How to study neural basis of behavior � Measure neural signaling in intact animals A. Tritonia and seapen � Implant a microcontroller in Tritonia brain � Tritonia is a model organism � Well studied neurophysiology � 500µm neurons; tolerant immune response � Work-in-progress B. Brain with implanted chip: Dorsal view Tritonia diomedea MEMS probe tip, amplifier brain visceral cavity tether memory battery microcontroller, A/D, cache Images courtesy James Beck & Russell Wyeth
An in-flight data recorder for insects � An autonomous microcontroller “in-the-loop” � Study neural basis of flight control Manduca Sext a or “hawk moth”
Participants � Chris Diorio and students from CSE � Karl Bohringer and students from EE (MEMS probes) � Tom Daniel and students from Zoology � Dennis Willows and students from Friday Harbor Labs � Funding from Packard, DoD MURI, NSF, DARPA, many others
Key point s � Don’t get hung up on t rying t o be a “pure science” � We are at t he cent er of everyt hing � There are incredible opport unit ies f or “peer t o peer” int ellect ual advancement � Beware of having a narrow view of what const it ut es comput er science
More recommend
