Popular in Engineering

Electronic Engineering

Control of a Wheelchair-Mounted Robotic Arm System

University of South Florida Redwan Alqasemi

Engineering

Spatiotemporal evaporating droplet dynamics on fomites enhances long term bacterial pathogenesis

Indian Institute of Science Saptarshi Basu

Civil Engineering

Suitability of excavated London clay as a supplementary cementitious material

ARUP Fragkoulis Kanavaris

Chemistry

Polyfluorene-Wrapped Carbon Nanotubes Decorated with Photocleavable Side-Chains

McMaster University Alex Adronov

Electronic Engineering

A Trusted and Privacy-Preserving Internet of Mobile Energy

Queensland University of Technology

Engineering

Water Security: Why We Need Global Solutions

Cardiff University

Engineering

Decision-Making Principles for Better Software Design Decisions

Swinburne University of Technology

Robust characterization of photonic integrated circuits

View more insights in Professional - Engineering - Electronic Engineering

View more figures from Monash University

View the page of Arthur Lowery

Read the Study

Next in Electronic Engineering

Motion-Informed Deep Learning for Brain MR Image Reconstruction Framework

Zhifeng Chen | Monash University

Control of a Wheelchair-Mounted Robotic Arm System

Redwan Alqasemi | University of South Florida

How can the dynamic performance of photonic integrated circuits (PICs) be robustly characterized to enable advanced applications?

Image courtesy of interviewee. December 4, 2024

Log-in or Sign-up to Faculti
Currently viewing this subject insight as a guest. You have insight(s) remaining for this month. Login to view 8000+ figures on the platform.

Generating Workspace...

Export as PDF

The speaker discusses their long-standing interest in photonic circuits, which consist of lasers and filters, comparing them to railway tracks. The goal is to create more stable circuits by integrating components onto a single silicon chip to minimize fluctuations caused by vibrations and temperature changes. They emphasize the need for reconfigurable chips that can be tuned post-production, akin to electronics field programmable gate arrays (FPGAs).

The presentation explains the challenges faced in achieving effective photonic circuits, including the limitations in the number of components due to light's properties and the need to handle optical interference for filtering purposes. They describe the concept of using finite impulse response filters to manipulate light paths, leading to frequency-dependent filtering.

The speaker also covers the importance of measuring phase response in optical systems and details the innovative use of Fourier transforms to achieve this. They note that they utilize a strong reference path to separate unwanted signals from the desired ones, thereby enhancing measurement accuracy. This video method allows for adaptive filtering, which can be applied in various contexts, such as telecommunications and image processing.

Additionally, the speaker mentions ongoing research in combining photonics with machine learning techniques, particularly through convolution processes, allowing for faster data processing. They conclude by addressing potential advancements in quantum computing, indicating that techniques developed for photonic circuits may prove useful in the quantum realm as well.

Language:

Export as PDF

Export as PDF
Here are five key points from the video regarding advancements and challenges in photonic circuits and their applications:

1. **Photonic Integrated Circuits**: The author discusses the development of photonic circuits that integrate components like lasers and filters onto common silicon substrates, which helps to reduce fluctuations due to environmental factors. This integration allows for stable and efficient operations without constant retuning.

2. **Reconfigurability and Tuning**: The author emphasizes the importance of creating reconfigurable or tuneable photonic chips. This capability, akin to field-programmable gate arrays (FPGAs) in electronics, allows researchers to adapt the functionalities of their photonic components post-production, enhancing flexibility and usability in various applications.

3. **Fourier Transforms and Signal Processing**: The discussion includes how Fourier transforms are used to analyze frequency responses in photonic circuits. By manipulating the paths of light, complex filtering is enabled, which has applications in neural networks and signal processing, particularly useful for tasks like image and audio processing.

4. **Adaptive Filtering**: The concept of adaptive filters in optical communications is presented, where the system can automatically adjust to optimize signal integrity. This includes components that compensate for dispersion in optical fibers, ensuring that signals maintain their quality over long distances.

5. **Challenges in Implementation**: The video outlines ongoing challenges in efficiently building and programming photonic circuits, including limitations on the number of components due to the nature of light. Additionally, it mentions competition with emerging technologies, such as quantum computing, indicating a dynamic environment in the field of photonics and optical communication.

Export as PDF
Here are some key quotes extracted from the video that encapsulate significant ideas and insights:

1. **Photonic Circuits and Their Fascination**:
"For many years I've been interested in photonic circuits, which are collections of lasers and filters, and they're rather like railway tracks that they all connect together, which is a fascination of mine."

2. **Challenges with Optical Systems**:
"the problem is you still get fluctuations that the fibers get vibrated or, the temperature changes and things like that, so you want it to sort of filter here and it moves across a bit and or goes all wiggly."

3. **Aim for Stability with Silicon**:
"so the aim is to put everything on a common piece of silicon or other material so that everything heats up together... you don't have to retune it all the time."

4. **Importance of Reconfigurability**:
"the idea really was to make a chip, which is reconfigurable, or at least tuneable. So once you've made it, you can finesse it and either make it do something else or move, move its response, its colors around."

5. **Analogy with Reconfigurable Electronics**:
"the flexibility of an FPGA... instead of making a custom chip... you can make processes that are good at images, processes that are good at audio processes, are good at computing or ai."

6. **Optical Interference as a Mechanism**:
"you get interference, optical interference. If you make those paths slightly different lengths, then the interference is frequency dependent."

7. **Filtering and Signal Processing**:
"operations like convolution, which is used a lot in neural networks, are really like a filter that you are, you're putting some signal through and it's going through these different paths and it's recombining."

8. **Measuring Complexity through Paths**:
"the more you have, the more complex the shape you can make in, in the filter."

9. **Using Fourier Transform for Measurements**:
"what we do, we use fourier transform in a slightly different way... we're using a thing that mathematically people wouldn't normally do, in the way we use the fourier transform."

10. **Application in Adaptive Filters**:
"instead of saying, I want you to be this shape, you say, I want the output of the optical link to be perfect... it looks at the output... and it’s a thing called an eye diagram."

11. **Convolution and Learning**:
"convolution is... matching images, matching sounds. Any sort of data is matching... the cerebellum at the back of your head here is a Big convolutional network, which is learning all the time."

12. **Quantum Computers as a Challenge**:
"there's also quantum computers... there's a big push in Australia to give a billion dollars."

These quotes capture the central themes of the discussion around photonic circuits, reconfigurability, optical interference, signal processing, and the application of convolution and adaptive filters, while also touching on the challenges presented by quantum computing.

X-TIMESTAMP-MAP=LOCAL:00:00:00.000,MPEGTS:90000

1
00:00:00.065 --> 00:00:03.605
For many years I've been interested in photonic circuits,

2
00:00:03.605 --> 00:00:06.205
which are collections of lasers and filters,

3
00:00:06.225 --> 00:00:08.405
and they're rather like railway tracks

4
00:00:08.405 --> 00:00:09.645
that they all connect together,

5
00:00:09.735 --> 00:00:11.485
which is a fascination of mine.

6
00:00:11.985 --> 00:00:15.165
Um, and the aim, really,

7
00:00:15.545 --> 00:00:17.725
you can build all this stuff on an optical bench,

8
00:00:17.775 --> 00:00:22.205
which is like a big bed, a big solid ion bed.

9
00:00:22.825 --> 00:00:26.405
And then the problem is you still get fluctuations

10
00:00:26.515 --> 00:00:29.205
that the fibers get vibrated

11
00:00:29.465 --> 00:00:32.605
or, the temperature changes and things like that,

12
00:00:32.985 --> 00:00:35.085
and the circuit goes unstable.

13
00:00:35.305 --> 00:00:37.205
So you want it to sort of filter here

14
00:00:37.305 --> 00:00:40.445
and it moves across a bit and or goes all wiggly.

15
00:00:40.985 --> 00:00:44.405
Um, so the aim is to put everything on a common piece

16
00:00:44.465 --> 00:00:47.445
of silicon or other material so

17
00:00:47.445 --> 00:00:49.045
that everything heats up together.

18
00:00:49.745 --> 00:00:51.885
Um, all the paths extend together,

19
00:00:52.465 --> 00:00:55.485
and that means that things don't shift.

20
00:00:56.065 --> 00:00:58.085
You don't have to retune it all the time.

21
00:00:59.145 --> 00:01:03.485
Um, so I had a grant, a Laureate fellowship, which,

22
00:01:04.065 --> 00:01:07.205
allowed me to experiment with different

23
00:01:07.735 --> 00:01:09.645
production techniques of chips

24
00:01:10.305 --> 00:01:13.485
and, we basically came to the conclusion that

25
00:01:14.045 --> 00:01:15.965
whatever you made, if you only made one of them,

26
00:01:16.915 --> 00:01:18.245
it's likely not to work.

27
00:01:18.585 --> 00:01:19.965
And I think that's generally

28
00:01:19.985 --> 00:01:21.765
for electronic chips as as well.

29
00:01:21.905 --> 00:01:25.645
So if you invest enough money, then obviously you get

30
00:01:25.825 --> 00:01:29.605
to a design, which you can use in telecommunication systems,

31
00:01:29.745 --> 00:01:34.285
but for researchers, that's just unfeasible really.

32
00:01:34.345 --> 00:01:35.725
The amount of money you have to spend

33
00:01:35.725 --> 00:01:38.285
and the amount of time that every time you do

34
00:01:38.285 --> 00:01:39.725
it takes six months.

35
00:01:40.625 --> 00:01:45.085
So the idea really was to, make a chip, which is

36
00:01:45.725 --> 00:01:47.605
reconfigurable, or at least tuneable.

37
00:01:48.145 --> 00:01:51.445
So once you've made it, you can finesse it

38
00:01:52.505 --> 00:01:54.645
and either make it do something else

39
00:01:54.825 --> 00:01:58.525
or move, move its response, its colors around,

40
00:01:59.985 --> 00:02:01.925
or just get it right.

41
00:02:02.425 --> 00:02:06.445
Uh, so that's the idea of photonic integrated.

42
00:02:06.465 --> 00:02:08.605
So it's one you put them on a chip that helps,

43
00:02:09.425 --> 00:02:12.885
but two, you really want

44
00:02:12.905 --> 00:02:15.965
to do post-production tuning,

45
00:02:16.785 --> 00:02:20.725
to really tune it up and make it what it needs to be.

46
00:02:21.025 --> 00:02:23.845
You know, like you have to tune a piano in a way.

47
00:02:24.365 --> 00:02:26.645
Reconfiguration in the electronics industry,

48
00:02:26.645 --> 00:02:29.525
there are things called field programmable gator

49
00:02:29.605 --> 00:02:30.845
arrays, FPGAs.

50
00:02:31.745 --> 00:02:34.765
And the idea here is instead of making a custom chip,

51
00:02:34.825 --> 00:02:37.725
that's everything's designed to exactly what you want,

52
00:02:38.265 --> 00:02:41.725
you have a sort of collection like Lego, if you like,

53
00:02:42.385 --> 00:02:46.765
and, you have a collection of logic gates and adders

54
00:02:46.765 --> 00:02:49.085
and multipliers, and then you have a sort

55
00:02:49.085 --> 00:02:50.925
of telephone exchange on top of this.

56
00:02:51.265 --> 00:02:53.005
So there's the, the basic logic,

57
00:02:53.145 --> 00:02:55.925
and then there's the wiring, if you like, on top of it.

58
00:02:56.225 --> 00:03:01.165
And by reprogramming the wiring, you can make anything out

59
00:03:01.165 --> 00:03:02.165
of an FPGA.

60
00:03:02.315 --> 00:03:04.845
Well, you can make processes that are good at images,

61
00:03:05.155 --> 00:03:08.085
processes that are good at audio processes,

62
00:03:08.145 --> 00:03:10.885
are good at computing or ai.

63
00:03:11.665 --> 00:03:15.965
And so the idea of photonic chips was to take

64
00:03:15.965 --> 00:03:20.085
that flexibility of an FPGA and try

65
00:03:20.085 --> 00:03:22.605
and bring it into the photonics world.

66
00:03:23.745 --> 00:03:25.245
Of course, you never really have

67
00:03:25.245 --> 00:03:27.725
as many components on a photonic chip

68
00:03:27.725 --> 00:03:30.245
because light doesn't go around corners very well,

69
00:03:30.305 --> 00:03:32.045
so they have to be a certain size.

70
00:03:32.145 --> 00:03:34.085
You can't put billions and billions of things.

71
00:03:34.505 --> 00:03:36.685
You could put hundreds and maybe thousands,

72
00:03:37.265 --> 00:03:40.085
but lighters is a very high frequency,

73
00:03:40.625 --> 00:03:42.645
and so you can turn it on

74
00:03:42.645 --> 00:03:45.285
and off very rapidly without really affecting it

75
00:03:45.665 --> 00:03:47.565
and push quite a lot of information

76
00:03:47.565 --> 00:03:49.965
through a photonic chip just as you can

77
00:03:49.965 --> 00:03:51.485
through an optical fiber.

78
00:03:53.205 --> 00:03:54.905
So fewer components,

79
00:03:55.445 --> 00:03:58.785
but greater bandwidth, should allow you

80
00:03:58.785 --> 00:03:59.985
to do something useful.

81
00:04:01.445 --> 00:04:06.225
So after these designs of various chips that didn't quite do

82
00:04:06.225 --> 00:04:09.865
what we wanted them to do, we came up with a design,

83
00:04:09.865 --> 00:04:11.145
which is a, a standard,

84
00:04:12.335 --> 00:04:14.705
it's called a finite impulse response filter

85
00:04:15.195 --> 00:04:18.865
where you have some inputs and you split the light up

86
00:04:19.245 --> 00:04:22.425
and you delay it a bit, and then you recombine it

87
00:04:22.425 --> 00:04:25.465
and it looks a bit like a railway shunting yard,

88
00:04:25.465 --> 00:04:27.665
putting wagons down different paths and things,

89
00:04:27.965 --> 00:04:29.625
but rather unlike wagons.

90
00:04:30.045 --> 00:04:32.625
If you put light down these different paths when it

91
00:04:32.945 --> 00:04:36.105
recombines at the far end, then you get interference,

92
00:04:36.215 --> 00:04:37.225
optical interference.

93
00:04:38.085 --> 00:04:40.505
If you make those paths slightly different lengths,

94
00:04:40.975 --> 00:04:43.745
then the interference is frequency dependent.

95
00:04:43.765 --> 00:04:46.925
So you start to get a filter that passes things

96
00:04:46.925 --> 00:04:48.005
and rejects things.

97
00:04:48.945 --> 00:04:52.685
Um, and you, a lot of operations like convolution,

98
00:04:52.695 --> 00:04:54.845
which is used a lot in neural networks,

99
00:04:55.225 --> 00:04:57.565
are really like a filter that you are,

100
00:04:57.705 --> 00:04:59.325
you're putting some signal through

101
00:04:59.385 --> 00:05:01.085
and it's going through these different paths

102
00:05:01.145 --> 00:05:03.965
and it's recombining, sometimes it's adding up,

103
00:05:03.965 --> 00:05:07.125
sometimes it's subtracting, that, that sort of thing.

104
00:05:07.505 --> 00:05:10.285
One of the challenges is when you build these chips,

105
00:05:10.625 --> 00:05:13.525
you can build this, we call it signal processing core,

106
00:05:14.095 --> 00:05:16.485
which is actually doing the maths if you like.

107
00:05:17.225 --> 00:05:20.685
Um, you connect it to some external equipment

108
00:05:20.685 --> 00:05:23.445
through optical fibers and you try and measure it.

109
00:05:23.545 --> 00:05:25.845
The signal processing core is this thing

110
00:05:25.845 --> 00:05:29.005
where you spit the light up and you recombine it

111
00:05:29.905 --> 00:05:32.925
and you make those paths slightly different lengths.

112
00:05:33.625 --> 00:05:35.725
And what that does is it gives you a ripple

113
00:05:35.865 --> 00:05:37.005
in the frequency domain.

114
00:05:37.105 --> 00:05:41.485
If you have just like two paths, we drew it here, you get,

115
00:05:41.785 --> 00:05:43.725
you know, a ripple like the blue ripple.

116
00:05:44.585 --> 00:05:47.925
Uh, if you have four paths, which is my fingers,

117
00:05:48.105 --> 00:05:50.805
1, 2, 3, 4 here, then each pair

118
00:05:50.825 --> 00:05:52.725
of those paths gives you a ripple.

119
00:05:52.865 --> 00:05:56.765
So one and two gives you the red Ripple two and three

120
00:05:56.775 --> 00:05:59.945
Gives you the, blue Ripple and three

121
00:05:59.945 --> 00:06:01.785
and four gives you the green ripple.

122
00:06:03.295 --> 00:06:05.555
Um, and you can combine them,

123
00:06:05.555 --> 00:06:08.355
but they'll always give you a ripple with the same,

124
00:06:08.585 --> 00:06:09.795
frequency or period.

125
00:06:10.615 --> 00:06:12.795
Uh, it might be smaller, it might be larger.

126
00:06:13.015 --> 00:06:16.275
The, the blue and the green cancel each other out here.

127
00:06:16.335 --> 00:06:20.155
So you'll basically be left with the red anyhow.

128
00:06:20.535 --> 00:06:24.635
Um, so that's what the signal processing core does.

129
00:06:24.655 --> 00:06:27.195
In fact, the light goes down all four paths

130
00:06:27.615 --> 00:06:29.115
and if you just take say one

131
00:06:29.115 --> 00:06:32.875
and four, that will have a, a ripple

132
00:06:33.145 --> 00:06:34.435
with a different period.

133
00:06:34.975 --> 00:06:37.035
And so you can start to get different shapes

134
00:06:37.575 --> 00:06:40.755
of in those ripples with more and more paths.

135
00:06:40.755 --> 00:06:44.115
The more you have, the more complex the shape you can make

136
00:06:44.695 --> 00:06:45.875
in, in the filter.

137
00:06:46.575 --> 00:06:51.035
So different colors traveling through it, here anyway,

138
00:06:52.335 --> 00:06:54.955
so the signal processing core is these four paths

139
00:06:55.695 --> 00:06:58.915
and you want some sort of reference,

140
00:06:58.915 --> 00:07:01.755
because if you just did that, compare that with that

141
00:07:01.755 --> 00:07:03.835
and that with that and that with that,

142
00:07:04.545 --> 00:07:06.315
then you get all muddled up.

143
00:07:06.315 --> 00:07:09.115
You don't know where the ripples are coming from, right?

144
00:07:09.575 --> 00:07:12.115
You don't know this is, you've got the sum

145
00:07:12.115 --> 00:07:14.435
of all the ripples, but you can't separate them out

146
00:07:14.435 --> 00:07:16.915
and go, Hey, you are the one to blame here.

147
00:07:17.015 --> 00:07:20.355
You are miss student. So the idea is you put a,

148
00:07:20.435 --> 00:07:22.595
a reference path and I'll, there's a pen here.

149
00:07:22.655 --> 00:07:26.635
So you, you put a path, which is a slightly different delay,

150
00:07:27.295 --> 00:07:31.915
and usually you have to make it far stronger so that it,

151
00:07:32.675 --> 00:07:35.095
and one get together to make a ripple it,

152
00:07:35.095 --> 00:07:36.775
and two, get together to make a ripple.

153
00:07:37.275 --> 00:07:40.495
It three gets together and make a ripple and it

154
00:07:40.495 --> 00:07:43.855
and four get to make a, a ripple.

155
00:07:45.035 --> 00:07:49.255
So it's, it's sort of mixing with 1, 2, 3, 4,

156
00:07:50.075 --> 00:07:51.895
to produce different ripples,

157
00:07:52.115 --> 00:07:56.055
but you've still got this problem of these guys interacting

158
00:07:56.055 --> 00:07:59.175
with themselves apart from this

159
00:07:59.635 --> 00:08:03.615
and that sort of muddles up your measurement a lot.

160
00:08:04.275 --> 00:08:06.815
So there's one solution as you make this so high

161
00:08:07.605 --> 00:08:10.055
that this mixing with one of these is far,

162
00:08:10.055 --> 00:08:12.415
far stronger than any two of these mixing together.

163
00:08:13.195 --> 00:08:16.775
But the other solution, is to move this over here.

164
00:08:16.925 --> 00:08:20.295
What that means is that,

165
00:08:20.765 --> 00:08:22.175
when you take a four year transform,

166
00:08:22.175 --> 00:08:25.255
which I explained later, it, it

167
00:08:25.895 --> 00:08:29.455
distinctly separates out all those different ripples.

168
00:08:29.835 --> 00:08:33.215
So all the ones that are between these sort

169
00:08:33.215 --> 00:08:34.255
of fall down here,

170
00:08:35.955 --> 00:08:40.055
but all the ones with this remain up here.

171
00:08:40.555 --> 00:08:43.735
So it all, the rubbish sort of is down in a,

172
00:08:43.925 --> 00:08:48.375
what usually is a big triangle like that of rubbish.

173
00:08:48.515 --> 00:08:50.175
So the solution is not make that stronger,

174
00:08:50.515 --> 00:08:54.135
but move it to the side rather simple, really.

175
00:08:54.715 --> 00:08:59.605
Uh, and that allows you to uniquely say,

176
00:08:59.825 --> 00:09:01.565
how much light is going through this path?

177
00:09:01.625 --> 00:09:04.005
How much is light is going through that path, et cetera.

178
00:09:04.665 --> 00:09:05.965
Uh, so the output

179
00:09:05.965 --> 00:09:08.965
of your computer tells you uniquely which one

180
00:09:08.965 --> 00:09:11.245
of these is needs tuning up, if you like.

181
00:09:12.465 --> 00:09:16.205
Um, rather than having a big muddle of, well,

182
00:09:16.235 --> 00:09:17.365
I've got this result,

183
00:09:17.425 --> 00:09:19.765
and I don't know whether it's this one misbehaving

184
00:09:19.785 --> 00:09:20.925
or this one misbehaving

185
00:09:21.025 --> 00:09:23.085
or some combination of the misbehaving.

186
00:09:23.315 --> 00:09:26.245
What we do, we use furier transform in a slightly different

187
00:09:26.245 --> 00:09:28.885
way, and I should explain what a furier transform does.

188
00:09:28.915 --> 00:09:31.285
It's generally if you want to go

189
00:09:31.285 --> 00:09:33.805
between a waveform like a musical,

190
00:09:34.675 --> 00:09:36.445
something coming outta your loud speaker

191
00:09:37.065 --> 00:09:38.805
and you want to know what notes are in it,

192
00:09:39.505 --> 00:09:41.045
you use a foria transform.

193
00:09:41.945 --> 00:09:44.165
If you want to know the beating of your heart,

194
00:09:44.235 --> 00:09:47.285
what frequency that is, you use a foria transform.

195
00:09:48.225 --> 00:09:52.845
Um, it's going between time and frequency.

196
00:09:53.145 --> 00:09:56.525
So it's telling you what notes are in a time varying

197
00:09:56.885 --> 00:09:59.405
waveform, which could be a voltage, it could be a pressure

198
00:10:00.275 --> 00:10:02.885
from a loudspeaker, could be anything.

199
00:10:04.505 --> 00:10:05.525
So what do we do?

200
00:10:05.525 --> 00:10:09.365
Well, it's quite easy to measure

201
00:10:10.065 --> 00:10:12.885
the response of a filter, so I should sort

202
00:10:12.885 --> 00:10:14.125
of draw one behind that.

203
00:10:14.915 --> 00:10:17.485
That might be the response of a filter.

204
00:10:17.705 --> 00:10:19.285
And this is frequency

205
00:10:19.345 --> 00:10:21.445
and that's how much gets through sort of thing.

206
00:10:22.225 --> 00:10:24.365
And you can do that with a tunable laser.

207
00:10:24.465 --> 00:10:27.365
So change the color of the laser, you measure

208
00:10:27.585 --> 00:10:29.645
how much power is coming outta the thing,

209
00:10:29.825 --> 00:10:31.325
and you plot a graph and

210
00:10:31.545 --> 00:10:33.645
that's measuring the intensity response.

211
00:10:34.425 --> 00:10:37.725
But it's very difficult to measure the phase response

212
00:10:37.725 --> 00:10:41.085
because light is sort of sign waves

213
00:10:41.085 --> 00:10:43.645
and cosine waves like these ones again.

214
00:10:44.505 --> 00:10:47.885
Um, and you want to know which is it, you know,

215
00:10:47.885 --> 00:10:50.605
is it shifted in time a little bit or not?

216
00:10:51.305 --> 00:10:53.165
Uh, and that's called phase.

217
00:10:53.705 --> 00:10:58.605
Um, for example, the, the, blue

218
00:10:58.625 --> 00:11:02.725
and the green are about 180 degrees apart and the red

219
00:11:02.725 --> 00:11:06.885
and the blue are maybe 90 degrees apart.

220
00:11:07.115 --> 00:11:10.085
That sort of thing and phase is really important

221
00:11:10.385 --> 00:11:11.645
for interference.

222
00:11:11.785 --> 00:11:14.765
So when you're putting light through all these paths

223
00:11:14.765 --> 00:11:17.805
and collecting it again, it's really important

224
00:11:18.835 --> 00:11:23.285
that little tiny delay of nanometers in one

225
00:11:23.285 --> 00:11:25.405
of the paths can give you a phase shift,

226
00:11:25.975 --> 00:11:28.205
which can turn the contribution on its head

227
00:11:28.225 --> 00:11:30.365
and make it negative rather than positive.

228
00:11:30.465 --> 00:11:35.005
So phase is really important in electronics and optics,

229
00:11:35.825 --> 00:11:36.965
but it's difficult to measure.

230
00:11:38.065 --> 00:11:41.045
Um, what we do with the Fourier transformers,

231
00:11:41.265 --> 00:11:44.245
we just measure the intensity with a photo dive

232
00:11:44.275 --> 00:11:47.645
that measures power, and we put this into the furier

233
00:11:47.965 --> 00:11:49.525
transform into its real input.

234
00:11:49.525 --> 00:11:52.285
So it's got a real and imaginary input,

235
00:11:53.375 --> 00:11:55.365
Which is another thing to explain maybe,

236
00:11:55.625 --> 00:11:57.525
but we just stick it into its real input

237
00:11:57.785 --> 00:12:02.325
and at the output we get both real and imaginary coming out.

238
00:12:02.545 --> 00:12:05.005
And from that we can work out phase

239
00:12:05.625 --> 00:12:06.845
and you're not supposed to do that.

240
00:12:07.305 --> 00:12:09.285
Uh, you're supposed to, you're supposed

241
00:12:09.285 --> 00:12:13.805
to put a real waveform in and get a real one out or reel

242
00:12:13.805 --> 00:12:15.925
and imaginary and get reel and imaginary out,

243
00:12:15.945 --> 00:12:18.795
but not just put reel in and get reel

244
00:12:18.795 --> 00:12:20.475
and imaginary out, which gives you face.

245
00:12:21.535 --> 00:12:24.785
And so it, it doesn't work perfectly.

246
00:12:25.445 --> 00:12:27.545
Um, it produces a lot of rubbish.

247
00:12:27.885 --> 00:12:32.825
Uh, but if I just draw this here, just as

248
00:12:32.845 --> 00:12:36.745
before where I was saying I need a new bit of paper,

249
00:12:37.845 --> 00:12:41.825
before I had these four paths

250
00:12:43.365 --> 00:12:47.145
and then I had a strong reference, which is like

251
00:12:47.145 --> 00:12:48.705
that woo.

252
00:12:49.445 --> 00:12:51.625
So that's covered up. So strong path,

253
00:12:51.625 --> 00:12:53.585
signal processing paths and reference.

254
00:12:54.415 --> 00:12:56.985
What happens when you use the furrow,

255
00:12:57.225 --> 00:12:59.865
transform it in this unusual way is

256
00:13:00.765 --> 00:13:03.425
all the rubbish got a different color.

257
00:13:04.325 --> 00:13:06.865
Um, falls,

258
00:13:07.515 --> 00:13:09.505
falls in this, this gap here.

259
00:13:09.605 --> 00:13:14.425
So it's, these are sort of bins if you like,

260
00:13:14.495 --> 00:13:17.385
like a graph and these are the ones you want.

261
00:13:17.725 --> 00:13:20.465
Ah, so by a using for a transform,

262
00:13:20.465 --> 00:13:22.385
which separates out all the delays,

263
00:13:23.085 --> 00:13:27.345
but b spacing this reference well away from the others,

264
00:13:28.325 --> 00:13:30.265
the rubbish falls in the gap.

265
00:13:31.645 --> 00:13:36.505
And that's sort of basically it really,

266
00:13:36.595 --> 00:13:38.065
we're, we're using a thing

267
00:13:38.065 --> 00:13:41.465
that mathematically people wouldn't normally do,

268
00:13:41.485 --> 00:13:44.665
in the way we use the furier transform with just one input.

269
00:13:45.445 --> 00:13:48.345
Uh, but we're getting around the imperfection

270
00:13:49.705 --> 00:13:52.765
by leaving a gap for the rubbish to fall in.

271
00:13:53.865 --> 00:13:55.485
And yeah, it works.

272
00:13:55.645 --> 00:13:57.325
I mean, mathematically it works as well,

273
00:13:57.385 --> 00:13:59.885
so you can write the equations and prove it's the truth.

274
00:13:59.885 --> 00:14:01.605
They've been doing a few things with it.

275
00:14:01.665 --> 00:14:05.125
The, the first one was to make a, a filter, which is,

276
00:14:05.125 --> 00:14:08.005
you know, something that selects different colors of light,

277
00:14:08.585 --> 00:14:10.565
that did what it was told.

278
00:14:11.105 --> 00:14:13.725
So we called it a dial up filter that you would say,

279
00:14:13.885 --> 00:14:17.605
I want this shape and it would control itself.

280
00:14:17.745 --> 00:14:20.925
Um, had little heaters on the guides on these guides

281
00:14:20.945 --> 00:14:22.845
to change their length slightly

282
00:14:22.905 --> 00:14:24.205
and change their phases slightly.

283
00:14:24.745 --> 00:14:25.925
It would control itself

284
00:14:25.945 --> 00:14:28.605
and give you what you wanted, which is good.

285
00:14:29.385 --> 00:14:33.125
Um, and then we did some fiber dispersion compensation.

286
00:14:33.345 --> 00:14:38.205
So in an optical fiber, blue goes faster than red normally,

287
00:14:38.945 --> 00:14:42.685
so you need some sort of handicap at the, the input

288
00:14:42.685 --> 00:14:46.165
or the output that says, Hey blue, you've come out first,

289
00:14:46.875 --> 00:14:48.365
I'll have to slow you down a bit

290
00:14:48.585 --> 00:14:50.165
and wait for red to come along.

291
00:14:50.945 --> 00:14:53.325
And you can do that with one of these filters

292
00:14:53.785 --> 00:14:55.005
and you can make it adaptive.

293
00:14:55.305 --> 00:14:59.165
So instead of saying, I want you to be this shape, you say,

294
00:14:59.765 --> 00:15:02.565
I want the output of the optical link to be perfect,

295
00:15:03.985 --> 00:15:07.165
and it looks at the output, the the ones and zeros

296
00:15:07.165 --> 00:15:08.605
and goes, those ones are a bit small

297
00:15:08.605 --> 00:15:10.325
and those zeros a bit large.

298
00:15:10.465 --> 00:15:12.845
And it, it's a thing called an eye diagram,

299
00:15:13.105 --> 00:15:15.405
and it sort of shrinks when things are bad

300
00:15:15.465 --> 00:15:17.405
and it gets bigger when things are good.

301
00:15:17.625 --> 00:15:22.605
So you look at how big that eye is, my I and,

302
00:15:23.785 --> 00:15:26.485
and you, you tune the filter.

303
00:15:27.065 --> 00:15:28.805
In fact, the computer tunes the filter.

304
00:15:29.745 --> 00:15:33.165
Uh, it's a bit like a neural network that it learns

305
00:15:34.635 --> 00:15:36.165
what it, it's looking at the eye

306
00:15:36.165 --> 00:15:39.805
and it's going hang on a bit, literally a bit.

307
00:15:39.805 --> 00:15:43.005
It's saying, I've got a one here and you transmit it a zero

308
00:15:43.025 --> 00:15:47.445
and every time you transmit a zero then a one you're wrong.

309
00:15:48.865 --> 00:15:52.205
So there must be something affecting from this one,

310
00:15:52.205 --> 00:15:53.205
affecting this one.

311
00:15:53.865 --> 00:15:58.605
So I'll take, you know, 20% of that and add it to this.

312
00:15:59.345 --> 00:16:02.605
Uh, so that's called an adaptive filter.

313
00:16:03.785 --> 00:16:05.685
Now that leads me on to neur networks,

314
00:16:05.685 --> 00:16:08.245
which are pretty much like adaptive filters,

315
00:16:08.385 --> 00:16:11.445
but they have many layers of them and they have many paths,

316
00:16:12.305 --> 00:16:15.605
but they, what they're trying to do all the time is say

317
00:16:16.195 --> 00:16:20.165
that little thing that weight, if, if I adjust that

318
00:16:21.125 --> 00:16:23.525
I get closer to learning the answer.

319
00:16:24.015 --> 00:16:26.765
There are some things that optics might be able to do better

320
00:16:26.765 --> 00:16:29.125
because you're pushing large amounts of information

321
00:16:29.125 --> 00:16:33.125
through the fiber, so you can learn at a quite a high rate.

322
00:16:34.305 --> 00:16:37.725
Um, and once you've learned, you just say stop learning

323
00:16:38.665 --> 00:16:41.045
or with chat GPT, it continues to learn

324
00:16:41.065 --> 00:16:44.845
and it gets stupider apparently, the more people

325
00:16:44.845 --> 00:16:46.965
that use it, the stupider it gets.

326
00:16:47.505 --> 00:16:51.805
Uh, so, that's aside there is

327
00:16:51.805 --> 00:16:56.005
that IEI aspect, there's optical communications,

328
00:16:57.025 --> 00:17:01.085
there's all sorts

329
00:17:01.085 --> 00:17:02.365
of things like looking at images.

330
00:17:02.815 --> 00:17:06.765
We've done some work, with my former,

331
00:17:07.085 --> 00:17:10.005
research fellow, Mike Sue, looking at images

332
00:17:10.225 --> 00:17:12.205
and pushing them through these systems.

333
00:17:13.065 --> 00:17:15.645
Um, it's basically doing convolution,

334
00:17:15.645 --> 00:17:16.885
which is pattern matching.

335
00:17:16.885 --> 00:17:19.045
Convolution is you've got something

336
00:17:19.505 --> 00:17:20.525
and you've got something else

337
00:17:20.545 --> 00:17:21.925
and you pass 'em through each other

338
00:17:22.505 --> 00:17:24.365
and you go, nearly there, nearly there.

339
00:17:24.425 --> 00:17:26.085
Oh no. Oh, that's a good match.

340
00:17:26.305 --> 00:17:28.005
Oh no, it's getting worse, it's getting worse.

341
00:17:28.505 --> 00:17:32.525
And you get this sort of triangle of improvement as it goes

342
00:17:32.525 --> 00:17:35.275
through, it gets better and worse.

343
00:17:36.255 --> 00:17:40.195
And convolutions used a lot for matching images,

344
00:17:41.495 --> 00:17:42.675
matching sounds.

345
00:17:43.895 --> 00:17:46.075
Any sort of data is matching.

346
00:17:46.075 --> 00:17:47.555
And I think the cerebellum at the back

347
00:17:47.555 --> 00:17:48.635
of your head here is a

348
00:17:48.985 --> 00:17:53.245
Big convolutional network, which is learning all the time

349
00:17:53.355 --> 00:17:55.125
that, you know, when I'm riding a bike,

350
00:17:55.455 --> 00:17:58.725
every time I'm on the pedals like this, every time I push

351
00:17:58.725 --> 00:18:00.845
that foot forward, I have to push

352
00:18:00.845 --> 00:18:02.605
that one forward a bit later.

353
00:18:03.235 --> 00:18:04.725
It's learning these patterns.

354
00:18:05.665 --> 00:18:09.485
So a lot of computation is, is based on that technique,

355
00:18:09.545 --> 00:18:12.045
and if you can push a lot of data through the,

356
00:18:12.785 --> 00:18:15.925
the convolutional neural network, you can train it quicker.

357
00:18:16.995 --> 00:18:19.925
That said, there are challenges,

358
00:18:20.265 --> 00:18:22.245
and talking of challenges, there's also,

359
00:18:22.875 --> 00:18:25.925
quantum computers that, there's a,

360
00:18:26.845 --> 00:18:30.045
a big push in Australia to give a billion dollars

361
00:18:30.185 --> 00:18:32.405
to a photonic quantum computer.

362
00:18:32.625 --> 00:18:34.965
And some of the circuits are pretty much like,

363
00:18:35.185 --> 00:18:36.245
you know, I've described.

364
00:18:36.945 --> 00:18:41.925
Um, and again, I'm hoping that some of these techniques

365
00:18:42.025 --> 00:18:45.765
for tuning up the circuits would be useful in that, in that

366
00:18:46.725 --> 00:18:46.845
scenario.

Advancements and Challenges in Photonic Circuits

00:00:00 - For many years I've been interested in photonic circuits, which are collections of lasers and filters, and they're rather like railway tracks that they all connect together, which is a fascination of mine. and the aim, really, you can build all this stuff on an optical bench, which is like a big bed, a big solid ion bed. And then the problem is you still get fluctuations that the fibers get vibrated or, the temperature changes and things like that,

00:00:35 - So you want it to sort of filter here and it moves across a bit and or goes all wiggly. so the aim is to put everything on a common piece of silicon or other material so that everything heats up together. all the paths extend together, and that means that things don't shift. You don't have to retune it all the time. so I had a grant, a Laureate fellowship, which,

00:01:07 - production techniques of chips and, we basically came to the conclusion that whatever you made, if you only made one of them, it's likely not to work. And I think that's generally for electronic chips as as well. So if you invest enough money, then obviously you get to a design, which you can use in telecommunication systems, but for researchers, that's just unfeasible really.

Advancements in Reconfigurable Photonic Chips

00:01:35 - and the amount of time that every time you do it takes six months. So the idea really was to, make a chip, which is reconfigurable, or at least tuneable. So once you've made it, you can finesse it and either make it do something else or move, move its response, its colors around, or just get it right. so that's the idea of photonic integrated.

00:02:09 - but two, you really want to do post-production tuning, to really tune it up and make it what it needs to be. You know, like you have to tune a piano in a way. Reconfiguration in the electronics industry, there are things called field programmable gator arrays, FPGAs. And the idea here is instead of making a custom chip, that's everything's designed to exactly what you want,

00:02:42 - and, you have a collection of logic gates and adders and multipliers, and then you have a sort of telephone exchange on top of this. So there's the, the basic logic, and then there's the wiring, if you like, on top of it. And by reprogramming the wiring, you can make anything out of an FPGA. Well, you can make processes that are good at images, processes that are good at audio processes, are good at computing or ai.

Integrating FPGA Flexibility into Photonic Circuits

00:03:15 - that flexibility of an FPGA and try and bring it into the photonics world. Of course, you never really have as many components on a photonic chip because light doesn't go around corners very well, so they have to be a certain size. You can't put billions and billions of things. You could put hundreds and maybe thousands, but lighters is a very high frequency, and so you can turn it on

00:03:45 - and push quite a lot of information through a photonic chip just as you can through an optical fiber. So fewer components, but greater bandwidth, should allow you to do something useful. So after these designs of various chips that didn't quite do what we wanted them to do, we came up with a design, which is a, a standard, it's called a finite impulse response filter

00:04:19 - and you delay it a bit, and then you recombine it and it looks a bit like a railway shunting yard, putting wagons down different paths and things, but rather unlike wagons. If you put light down these different paths when it recombines at the far end, then you get interference, optical interference. If you make those paths slightly different lengths, then the interference is frequency dependent. So you start to get a filter that passes things

Signal Processing in Photonic Circuits: Convolution and Frequency Ripple Analysis

00:04:48 - and you, a lot of operations like convolution, which is used a lot in neural networks, are really like a filter that you are, you're putting some signal through and it's going through these different paths and it's recombining, sometimes it's adding up, sometimes it's subtracting, that, that sort of thing. One of the challenges is when you build these chips, you can build this, we call it signal processing core, which is actually doing the maths if you like.

00:05:20 - through optical fibers and you try and measure it. The signal processing core is this thing where you spit the light up and you recombine it and you make those paths slightly different lengths. And what that does is it gives you a ripple in the frequency domain. If you have just like two paths, we drew it here, you get, you know, a ripple like the blue ripple. if you have four paths, which is my fingers,

00:05:50 - of those paths gives you a ripple. So one and two gives you the red Ripple two and three Gives you the, blue Ripple and three and four gives you the green ripple. and you can combine them, but they'll always give you a ripple with the same, frequency or period. it might be smaller, it might be larger. The, the blue and the green cancel each other out here. So you'll basically be left with the red anyhow.

Manipulating Light Paths in Photonic Filters

00:06:24 - In fact, the light goes down all four paths and if you just take say one and four, that will have a, a ripple with a different period. And so you can start to get different shapes of in those ripples with more and more paths. The more you have, the more complex the shape you can make in, in the filter. So different colors traveling through it, here anyway,

00:06:55 - and you want some sort of reference, because if you just did that, compare that with that and that with that and that with that, then you get all muddled up. You don't know where the ripples are coming from, right? You don't know this is, you've got the sum of all the ripples, but you can't separate them out and go, Hey, you are the one to blame here. You are miss student. So the idea is you put a, a reference path and I'll, there's a pen here.

00:07:27 - and usually you have to make it far stronger so that it, and one get together to make a ripple it, and two, get together to make a ripple. It three gets together and make a ripple and it and four get to make a, a ripple. So it's, it's sort of mixing with 1, 2, 3, 4, to produce different ripples, but you've still got this problem of these guys interacting

Optimizing Measurements with Fourier Transform Techniques

00:07:59 - and that sort of muddles up your measurement a lot. So there's one solution as you make this so high that this mixing with one of these is far, far stronger than any two of these mixing together. But the other solution, is to move this over here. What that means is that, when you take a four year transform, which I explained later, it, it distinctly separates out all those different ripples.

00:08:33 - of fall down here, but all the ones with this remain up here. So it all, the rubbish sort of is down in a, what usually is a big triangle like that of rubbish. So the solution is not make that stronger, but move it to the side rather simple, really. and that allows you to uniquely say, how much light is going through this path?

00:09:04 - so the output of your computer tells you uniquely which one of these is needs tuning up, if you like. rather than having a big muddle of, well, I've got this result, and I don't know whether it's this one misbehaving or this one misbehaving or some combination of the misbehaving. What we do, we use furier transform in a slightly different way, and I should explain what a furier transform does. It's generally if you want to go between a waveform like a musical,

Understanding Fourier Transforms and Phase Response in Signal Processing

00:09:37 - and you want to know what notes are in it, you use a foria transform. If you want to know the beating of your heart, what frequency that is, you use a foria transform. it's going between time and frequency. So it's telling you what notes are in a time varying waveform, which could be a voltage, it could be a pressure from a loudspeaker, could be anything. So what do we do?

00:10:10 - the response of a filter, so I should sort of draw one behind that. That might be the response of a filter. And this is frequency and that's how much gets through sort of thing. And you can do that with a tunable laser. So change the color of the laser, you measure how much power is coming outta the thing, and you plot a graph and that's measuring the intensity response. But it's very difficult to measure the phase response

00:10:41 - and cosine waves like these ones again. and you want to know which is it, you know, is it shifted in time a little bit or not? and that's called phase. for example, the, the, blue and the green are about 180 degrees apart and the red and the blue are maybe 90 degrees apart. That sort of thing and phase is really important

Measuring Phase Shifts in Photonic Circuits

00:11:11 - So when you're putting light through all these paths and collecting it again, it's really important that little tiny delay of nanometers in one of the paths can give you a phase shift, which can turn the contribution on its head and make it negative rather than positive. So phase is really important in electronics and optics, but it's difficult to measure. what we do with the Fourier transformers,

00:11:44 - that measures power, and we put this into the furier transform into its real input. So it's got a real and imaginary input, Which is another thing to explain maybe, but we just stick it into its real input and at the output we get both real and imaginary coming out. And from that we can work out phase and you're not supposed to do that. you're supposed to, you're supposed to put a real waveform in and get a real one out or reel

00:12:15 - but not just put reel in and get reel and imaginary out, which gives you face. And so it, it doesn't work perfectly. it produces a lot of rubbish. but if I just draw this here, just as before where I was saying I need a new bit of paper, before I had these four paths and then I had a strong reference, which is like

Innovative Use of Fourier Transforms in Signal Processing

00:12:49 - So that's covered up. So strong path, signal processing paths and reference. What happens when you use the furrow, transform it in this unusual way is all the rubbish got a different color. falls, falls in this, this gap here. So it's, these are sort of bins if you like, like a graph and these are the ones you want.

00:13:20 - which separates out all the delays, but b spacing this reference well away from the others, the rubbish falls in the gap. And that's sort of basically it really, we're, we're using a thing that mathematically people wouldn't normally do, in the way we use the furier transform with just one input. but we're getting around the imperfection

00:13:53 - And yeah, it works. I mean, mathematically it works as well, so you can write the equations and prove it's the truth. They've been doing a few things with it. The, the first one was to make a, a filter, which is, you know, something that selects different colors of light, that did what it was told. So we called it a dial up filter that you would say, I want this shape and it would control itself. had little heaters on the guides on these guides

Adaptive Fiber Dispersion Compensation and Self-Tuning Filters

00:14:22 - and change their phases slightly. It would control itself and give you what you wanted, which is good. and then we did some fiber dispersion compensation. So in an optical fiber, blue goes faster than red normally, so you need some sort of handicap at the, the input or the output that says, Hey blue, you've come out first, I'll have to slow you down a bit and wait for red to come along.

00:14:53 - and you can make it adaptive. So instead of saying, I want you to be this shape, you say, I want the output of the optical link to be perfect, and it looks at the output, the the ones and zeros and goes, those ones are a bit small and those zeros a bit large. And it, it's a thing called an eye diagram, and it sort of shrinks when things are bad and it gets bigger when things are good. So you look at how big that eye is, my I and,

00:15:27 - In fact, the computer tunes the filter. it's a bit like a neural network that it learns what it, it's looking at the eye and it's going hang on a bit, literally a bit. It's saying, I've got a one here and you transmit it a zero and every time you transmit a zero then a one you're wrong. So there must be something affecting from this one, affecting this one.

Adaptive Filters and Neural Networks: Learning Through Convolution

00:15:59 - so that's called an adaptive filter. Now that leads me on to neur networks, which are pretty much like adaptive filters, but they have many layers of them and they have many paths, but they, what they're trying to do all the time is say that little thing that weight, if, if I adjust that I get closer to learning the answer.

00:16:26 - because you're pushing large amounts of information through the fiber, so you can learn at a quite a high rate. and once you've learned, you just say stop learning or with chat GPT, it continues to learn and it gets stupider apparently, the more people that use it, the stupider it gets. so, that's aside there is that IEI aspect, there's optical communications,

00:17:01 - of things like looking at images. We've done some work, with my former, research fellow, Mike Sue, looking at images and pushing them through these systems. it's basically doing convolution, which is pattern matching. Convolution is you've got something and you've got something else and you pass 'em through each other and you go, nearly there, nearly there. no. that's a good match. no, it's getting worse, it's getting worse.

Convolutional Networks and Adaptive Learning

00:17:32 - through, it gets better and worse. And convolutions used a lot for matching images, matching sounds. Any sort of data is matching. And I think the cerebellum at the back of your head here is a Big convolutional network, which is learning all the time that, you know, when I'm riding a bike, every time I'm on the pedals like this, every time I push

00:18:00 - that one forward a bit later. It's learning these patterns. So a lot of computation is, is based on that technique, and if you can push a lot of data through the, the convolutional neural network, you can train it quicker. That said, there are challenges, and talking of challenges, there's also, quantum computers that, there's a, a big push in Australia to give a billion dollars

00:18:46 - And some of the circuits are pretty much like, you know, I've described. and again, I'm hoping that some of these techniques for tuning up the circuits would be useful in that, in that scenario.

Faculti AI can make mistakes

Research

Policy

Analysis