Page 1
PHOTONICS RESEARCH GROUP 1
PHOTONICS RESEARCH GROUP
Photonic reservoir computing using silicon chips
Kristof Vandoorne, Pauline Mechet, Martin Fiers, Thomas Van Vaerenbergh, Bendix Schneider, Andrew Katumba, Floris Laporte, David Verstraeten, Benjamin Schrauwen, Joni Dambre and Peter Bienstman
Page 2
PHOTONICS RESEARCH GROUP 2
THE BLACK BOX
Page 3
PHOTONICS RESEARCH GROUP 3
What can this chip do?
Page 4
PHOTONICS RESEARCH GROUP 4
Several things!
• Do arbitrary boolean calculations with memory on a bitstream
• Recognise arbitrary 5-bit headers at 12.5 Gbps
• Perform speech recognition of isolated digits
• Does not consume any active power
• Easily upscalable to higher speeds
Page 5
PHOTONICS RESEARCH GROUP 5
How does it do it?
Using “Reservoir computing”, a brain-inspired technique to solve pattern recognition problems in a fast and power-efficient way
Page 6
PHOTONICS RESEARCH GROUP 6
WHAT IS RESERVOIR COMPUTING?
Page 7
PHOTONICS RESEARCH GROUP 7
What is reservoir computing?
• From field of machine learning (2002)
• Related to neural networks
• So far mainly in software
• Very successful:• Better than state-of-the-art digit recognition
• Speech recognition
• Robot control
• …
Page 8
PHOTONICS RESEARCH GROUP 8
Reservoir Readout
Reservoir computing
Don’t train the neural network, only train the linear readout
Page 9
PHOTONICS RESEARCH GROUP 9
reservoir state
readout
reservoir
nothing pebbles gritpebbles
grit
A hardware implementation…
Page 10
PHOTONICS RESEARCH GROUP 10
*
**
**
*
*
●●
● ●
●
●
x
y
z'
*
** *
***x'
y'
To higher order space
●●● ●
●●
Why does it work?
Page 11
PHOTONICS RESEARCH GROUP 11
PHOTONIC RESERVOIR COMPUTING
Page 12
PHOTONICS RESEARCH GROUP 12
Photonics
Photonic reservoirs
Page 13
PHOTONICS RESEARCH GROUP 13
• Faster• More power efficient• Richer dynamics in nodes• Light has a phase
Why photonics?
Page 14
PHOTONICS RESEARCH GROUP 14
OPTICAL AMPLIFIER NETWORKSThe very beginning…
Page 15
PHOTONICS RESEARCH GROUP 15
Looks like tanh, but positive signals only
Output
Use SOAs as neurons
Page 16
PHOTONICS RESEARCH GROUP 16
The gain in the SOA model is dependent on the input power and its own history
SOA model
Page 17
PHOTONICS RESEARCH GROUP 17
81 SOAs
Swirl topology
Page 18
PHOTONICS RESEARCH GROUP 18
5 female speakers, saying
10 times the same 10 digits,
ranging from zero to nine
Speech corpus
Page 19
PHOTONICS RESEARCH GROUP 19
• dynamics of light signal should be on time scale of SOA dynamics and chip delays
• convert 1 sec speech to 1 ns light signal
• 9 orders of magnitude upconversion
Time scales
Page 20
PHOTONICS RESEARCH GROUP 20
Word error rate
Page 21
PHOTONICS RESEARCH GROUP 21
Optimal delay
75 ps
187.5 ps
312.5 ps
Page 22
PHOTONICS RESEARCH GROUP 22
Absolute minimum
(phase controlled)Minimum
(phase averaged)
Reducing 2D plots to single number
Page 23
PHOTONICS RESEARCH GROUP 23
Controlling the phase offers clear advantage
Page 24
PHOTONICS RESEARCH GROUP 24
PASSIVE SILICON RESERVOIRSThe next step…
Page 25
PHOTONICS RESEARCH GROUP 25
What happens if you remove the SOAs?
Page 26
PHOTONICS RESEARCH GROUP 26
Passive Silicon reservoir
• silicon photonics: mature technology
• nodes become simple splitters/combiners
• non-linearity in readout suffices
• no need for amplifiers which consume power
• no longer limited by timescale of non-linearity
Vandoorne et al, Nature Comms, 5, 3541, 2014
Page 27
PHOTONICS RESEARCH GROUP 27
NL coming from the detector suffices!
Speech task: passive reservoirs (no amplifiers)
Page 28
PHOTONICS RESEARCH GROUP 28
16 node swirl network where 11 nodes could be measured from 1 input
Page 29
PHOTONICS RESEARCH GROUP 29
The input: 11136 bits modulated at 1531 nm with speeds between 125Mbit/s and 12.5Gbit/s
Page 30
PHOTONICS RESEARCH GROUP 30
First task: desired output should be the XOR of every bit with the previous bit.
Hard task in machine learning (non-linear!)
Page 31
PHOTONICS RESEARCH GROUP 31
Measurements and simulations for the XOR task correspond
Page 32
PHOTONICS RESEARCH GROUP 32
The XOR task can be solved at different speeds and different bit combinations
Page 33
PHOTONICS RESEARCH GROUP 33
Other Boolean tasks can be solved as well (with the same reservoir states)
Page 34
PHOTONICS RESEARCH GROUP 34
Header recognition
Page 35
PHOTONICS RESEARCH GROUP 35
Advantages
• Scalability: • Note that we spent a lot of effort to slow down the signal!
• Easily scalable to higher speeds by shortening the delays
• No active power consumption on chip
• Same generic chip can be used for• digital tasks (simulation confirmed by experiment)
• analog tasks (theory only, no suitable equipment)
Page 36
PHOTONICS RESEARCH GROUP - CONFIDENTIAL 36
APPLICATIONS
Page 37
PHOTONICS RESEARCH GROUP 37
Telecom task: non-linear equalization of optical links
Page 38
PHOTONICS RESEARCH GROUP 38
Signal Equalization: Results….
Up to 200 km below FEC Limit
Metro Links
Equalization results with passive SOI chip
Page 39
PHOTONICS RESEARCH GROUP 39
Scaling this up
• PhResCo: recently started H2020 European project (KULeuven, IBM, UGent, Supelec, IHP)
• Integrated readout on chip:
out…
…
…
…in
reservoir readout
Page 40
PHOTONICS RESEARCH GROUP 40
First design: comparing 3 different technologies
2 x 9 Reservoir
BTO Test Structures
Si Readout BTO Readout
VO2 Readout
VO2 Test Structures
Page 41
PHOTONICS RESEARCH GROUP 41
Conclusions
Neuromorphic computing
is interesting new paradigm
for photonics information processing
Page 42
PHOTONICS RESEARCH GROUP 42
Flow cytometry
http://www.lifetechnologies.com
Page 43
PHOTONICS RESEARCH GROUP 43
Imec cell sorter
Integrated micro-fluidic
channels
On-chip high speed
cell sorting
On-chip Fast
high-resolution microscopy
Page 44
PHOTONICS RESEARCH GROUP 44
Computational complexity
➢ Complex convolution or sequence of 2D FFTs
➢ 512x512 pixels/image
➢ 1M cells/sec
➢ 48.8M Flops for reconstruction
➢ ~ 60 TFlops/sec including classification
http://www.top500.org/
# Site System Cores Perf.ormance[TF/sec]
Power [kW]
482 AutomotiveUnited States
IBM Flex System x240, Xeon E5-2670 8C 2.600GHz, Infiniband FDR IBM
8,336 157.7 181
Page 45
PHOTONICS RESEARCH GROUP 45
Algorithm Methods
classificationfeature
selectionnumerical
reconstruction
Lymphocytes
Monocytes
Granulocytes
Page 46
PHOTONICS RESEARCH GROUP 46
Real experimental data
k
1.39
1.37
1.34E
Direction of flow
Incident plane wave Scattered wave +
Direct wave
Detector plane
Microfluidic flow
chamber
Page 47
PHOTONICS RESEARCH GROUP 47
Neural network - pipeline
.
.
.
.
.
.
Input Layer Hidden Layer Output Layer
ANN: < 200 GigaFlop/sec !
Page 48
PHOTONICS RESEARCH GROUP 48
Three-part WBC classification Results
• Dataset of ~7500 non-purified WBC:
Granulocytes (59.8%),
Lymphocytes (34.6%),
Monocytes (5.6%)
• Use of 10 random folds for cross-
validating (CV) the results
• Adding noise to weights at fixed SNR
Page 49
PHOTONICS RESEARCH GROUP 49
Purified monocyte/granulocyte classification
Averaged classification results with increasing signal-to-noise ratio (from left to right: 30dB, 10 dB, 3 dB)
Class 1 = monocytesClass 2 = granulocytes
Page 50
PHOTONICS RESEARCH GROUP 50
Towards a hardware solution
.
.
.
.
.
.
Page 51
PHOTONICS RESEARCH GROUP 51
Conclusions
Neuromorphic computing
is interesting new paradigm
for photonics information processing
Page 52
PHOTONICS RESEARCH GROUP 52
EXCITABLE SILICON RINGS
Page 53
PHOTONICS RESEARCH GROUP 53
Building a photonic spiking neuron
= ?
Page 54
PHOTONICS RESEARCH GROUP 54
Research question
• People have seen excitability in photonics before, but never cascaded it on chip
• Can we cascade excitability on-chip using ring-resonator neurons?
Page 55
PHOTONICS RESEARCH GROUP 55
Thermo-optic effect causes redshift
Light circulation in ring resonance dip/peak
Heating of the ring redshiftT
ire
Page 56
PHOTONICS RESEARCH GROUP 56
Self-heating causes bistability
Light circulation in ring resonance dip/peak
Heating of the ring redshiftT
ire
Page 57
PHOTONICS RESEARCH GROUP 57
Free carriers cause blueshift
Light circulation in ring resonance dip/peak
Free carriers blueshift
ire
N
Page 58
PHOTONICS RESEARCH GROUP 58
Combination free carrier and thermal effect can cause self-pulsation
Light circulation in ring ~ ps
Cooling of the ring ~ 100 ns
Free carriers ~ ns
ire
N
T
1
2
3
Page 59
PHOTONICS RESEARCH GROUP 59
Simulations: bistability and self-pulsation
Q 6.25 104
25 pm62 pm
R 4 μm
dB3
r −
Page 60
PHOTONICS RESEARCH GROUP 60
Simulation: excitability
Wavelength and input power ‘near’ self-pulsation...
Page 61
PHOTONICS RESEARCH GROUP 61
Simulation: cascadability
Page 62
PHOTONICS RESEARCH GROUP 62
Experiment: self-pulsation
Page 63
PHOTONICS RESEARCH GROUP 63
Experiment: excitability
Pulses excited by external trigger signal:
Page 64
PHOTONICS RESEARCH GROUP 64
Experiment: cascadability
Page 65
PHOTONICS RESEARCH GROUP 65
Cascading rings = creating a delay line
…
t
t
t
t
tt
t
Page 66
PHOTONICS RESEARCH GROUP 66
Cascading rings = creating a delay line
…
t
t
t
t
tt
t
Max ~ 9-10 rings
t
…
Page 67
PHOTONICS RESEARCH GROUP 67
10 rings result in a ~200 ns delay of a 15-20 ns pulse
Page 68
PHOTONICS RESEARCH GROUP 68
10 rings result in a ~200 ns delay of a 15-20 ns pulse
Page 69
PHOTONICS RESEARCH GROUP 69
Making a loop => spike encoded memory/clock
…
t
t t
…
If delay > internal timescale neuron
=> Excitation loops through rings
Page 70
PHOTONICS RESEARCH GROUP 70
The concept works! (loop from ring 2-8)
Page 71
PHOTONICS RESEARCH GROUP 71
Conclusions
Neuromorphic computing
is interesting new paradigm
for photonics information processing