22
Internet of Smart Things where machine learning meets embedded systems Advanced Research I ntelligent E mbedded S ystems Politecnico di Milano Prof. Manuel Roveri Dipartimento di Elettronica, Informazione e Bioingegneria Politecnico di Milano

Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Internet  of  Smart  Thingswhere machine learning meets embedded systems

Adv

ance

d Re

sear

ch Intelligent Embedded Systems

Politecnico di Milano

Prof.  Manuel  RoveriDipartimento  di  Elettronica,  Informazione  e  Bioingegneria  

Politecnico  di  Milano  

Page 2: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

“Stand  on  the  shoulders  of  Giants”

Alan Turing (1912-1954)Claude  Shannon (1916-2001)

Computer  science  and  AI:  the  rise  of  both  disciplines

Page 3: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Computer  science  and  AI:  the  rise  of  both  disciplines

Claude  Shannon (1916-2001)

A

B

C

“A  Symbolic  Analysis  of  Relay  and  Switching  Circuits”,  Master's  degree  thesis,  Massachusetts  Institute  of  Technology,  1937

Page 4: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Computer  science  and  AI:  the  rise  of  both  disciplines

Claude  Shannon (1916-2001)

A=“Fever”

B=“Flu”

C=“Night”

“A  Symbolic  Analysis  of  Relay  and  Switching  Circuits”,  Master's  degree  thesis,  Massachusetts  Institute  of  Technology,  1937

“ill”

“Stay  at  home”

Page 5: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

TURING  TEST

Computer  science  and  AI:  the  rise  of  both  disciplines

Alan Turing (1912-1954)

Garry  Kasparov  (2012)  playing  Turing’s  Chess  (1948)

A  test  for  a  machine  to  be  called  "intelligent”

Page 6: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

TURING  TEST

Computer  science  and  AI:  the  rise  of  both  disciplines

Alan Turing (1912-1954)

Garry  Kasparov  (2012)  playing  Turing’s  Chess  (1948)

A  test  for  a  machine  to  be  called  "intelligent”

Page 7: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Generations  and  seasons

1945

Eniac

Univac

1960 1970 1980 Today

Transistor Integrated  Circuits

Microprocessors

StatisitcalMethods

Pioneering  works  in  the  field  of  AI

Bayesian  Inference

Neural  Networks

K  Nearest  Neighbour

SVMDeep  Learning

BackProp

Page 8: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Generations  and  seasons:  the  evolution   of  computation  and  memory  

1945

Eniac

Univac

1960 1970 1980 Today

Transistor Integrated  Circuits

Microprocessors

StatisitcalMethods

Pioneering  works  in  the  field  of  AI

Bayesian  Inference

Neural  Networks

K  Nearest  Neighbour

SVMDeep  Learning

BackProp

Computational  PowerAvailable  Memory

Computational  needsMemory  Requirements

Page 9: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

20x  – 50x 100x  – 1000x

50x  –100x

1Kx  –10Kx

GPU,  FPGANeural  HW

SupercomputersHPC

RAM

ComputationSpeedup

PC

Embedded  PCs

Embedded  Systems

0.0001x  – 0.0005x

0.1x  –0.5x

0.01x–0.05x

AI  and  Technology

Page 10: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

20x  – 50x 100x  – 1000x

50x  –100x

1Kx  –10Kx

GPUs,Neural  HW

Supercomputers(Systems  of  GPUSs  or  Neural  HWs)

RAM

ComputationSpeedup

Embedded  PCs

Embedded  Systems

0.0001x  – 0.0005x

0.1x  –0.5x

0.01x–0.05x

AI  and  Technology

Intelligent Cyber-Physical Systems

Deep Learning

GPU,  FPGANeural  HW

SupercomputersHPC

PC

Page 11: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Intelligent  Internet-­of-­Things  and  Cyber-­Physical  SystemsCyber

Domain

Physical  

Domain

Self-­awareness

Self-­Diagnosis

Reliability

Self-­healing

Adaptation

Intelligent  Processing  

ofPhysical  Sensing

Cognitive  Mechanisms  for  Actuation  and  Control

Intelligent  IoT and  Cyber-­Physical  Systems

Page 12: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

POLIMI  and  STMicroelectronics:Designing   Intelligent  Cyber-­Physical  Systems

Page 13: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

The  Intelligent   Embedded  Sensors:  learning  Recurrent  Neural  Networks  (ESNs)

Trained  on  a  Coordinatorof

the  CPS

Trained  on  the  Intelligent  Embedded  Sensor

Page 14: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

The  Intelligent   Embedded  Sensors:  robustness  mechanisms  to  shield  removal

Self-­‐ability  to  manage  the  removal  and  the  insertion  of  the  sensor  shield  board   from  the  STM32  main  board

Page 15: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

The  Coordinator:  dependency-­graph   learning  and  network  management

Page 16: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

The  Server:  data  analysis,  processing  and  visualization

http://131.175.156.3:5984/stdashboard/_design/ST-­DASHBOARD/main2.html

Page 17: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

The  Testbed:   Intelligent  Monitoring  of  Datacenters

Page 18: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

What  about  Deep  Learning?  

20x  – 50x 100x  – 1000x

50x  –100x

1Kx  –10Kx

GPUs,Neural  HW

Supercomputers(Systems  of  GPUSs  or  Neural  HWs)

RAM

ComputationSpeedup

PC

Embedded  PC

Embedded  Systems

0.05x  – 0.1x0.0001x  – 0.0005x

0.1x  –0.5x

0.01x–0.05x

Deep Learning

Page 19: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Deep Learning

What  about  Deep  Learning?  

20x  – 50x 100x  – 1000x

50x  –100x

1Kx  –10Kx

RAM

ComputationSpeedup

PC

Embedded  Systems

0.05x  – 0.1x0.0001x  – 0.0005x

0.1x  –0.5x

0.01x–0.05x

Embedded  Systems

Page 20: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

How  to  meet  Embedded  Systems  with  Deep  Learning?

20x  – 50x 100x  – 1000x

50x  –100x

1Kx  –10Kx

RAM

ComputationSpeedup

PC

Embedded  Systems

Approximate  Computing

0.05x  – 0.1x0.0001x  – 0.0005x

0.1x  –0.5x

0.01x–0.05x

Embedded  System  Code  Optimization

Re-­design  of  the  CNN  architecture

Page 21: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Progetto Regione Lombardia -­ Smart  Living  (2018-­2019)

Sistema Intelligente per  il M onitoraggio e  la  Predizione della Solidità Strutturale di  edifici e  infrastrutture e  per  la  pianificazione dell’intervento -­ SIMPSS  

Page 22: Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers (Systems&ofGPUSs& or&Neural&HWs) RAM Computation Speedup Embedded& PCs Embedded&

Thank  you  for  the  attention!