Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers...
Transcript of Presentazione ASTday 2018 Roveri PDF · 2018. 2. 28. · 1K x – x GPUs, Neural&HW Supercomputers...
![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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/1.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/2.jpg)
“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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/3.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/4.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/5.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/6.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/7.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/8.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/9.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/10.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/11.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/12.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/13.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/14.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/15.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/16.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/17.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/18.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/19.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/20.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/21.jpg)
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&](https://reader035.fdocuments.us/reader035/viewer/2022071214/6042e0c6221462326e63c382/html5/thumbnails/22.jpg)
Thank you for the attention!