Quantum Dynamics vs Classical Physics — Entangled Health Experiments

WEBVTT

00:00:30.000 --> 00:00:35.400
and I'll play you their script pause when it's sharp pause when it's slipped and

00:00:36.420 --> 00:00:40.460
You can hear what's a wife's and if you've been with entangled health for a while

00:00:40.800 --> 00:00:47.360
I'm especially curious how this lands whether this still feels like entangled health to you and

00:00:48.559 --> 00:00:51.879
If it doesn't I want to know imagine taking a

00:00:51.879 --> 00:00:52.580
ahh

00:00:52.900 --> 00:00:57.980
massive artificial intelligence, like a neural network with hundreds of processors.

00:00:58.460 --> 00:01:01.720
-Right. -The kind that consumes ridiculous amounts of power,

00:01:02.040 --> 00:01:04.900
requiring a whole server farm just to keep the chips from melting.

00:01:05.240 --> 00:01:08.500
-Oh, yeah, the energy footprint on those as well. It's massive. -Exactly.

00:01:09.120 --> 00:01:15.640
And then imagine replacing that entire sprawling architecture with just, and I mean literally, five individual atoms.

00:01:16.340 --> 00:01:17.940
-It sounds completely absurd. -It does.

00:01:19.820 --> 00:01:22.000
Wait, wait. Let's listen to that again.

00:01:22.500 --> 00:01:29.300
Yeah, sprawling architecture with just and I mean literally five individual atoms. I mean come on this is completely absurd

00:01:30.100 --> 00:01:33.479
Yes, but this is indeed not the style I wanted to have

00:01:35.580 --> 00:01:38.580
Let's continue listening. I think they have a bit of too much hype here, but

00:01:38.740 --> 00:01:39.640
Um...

00:01:39.700 --> 00:01:40.880
Let's go on.

00:01:56.000 --> 00:01:57.900
100 to 500 nodes.

00:01:58.220 --> 00:02:02.260
Yeah, it's one of those statistics that it forces you to go back and reread the paper.

00:02:05.720 --> 00:02:06.820
Okay, let's pause right here.

00:02:07.140 --> 00:02:08.560
Forces you to reread the paper.

00:02:08.759 --> 00:02:10.060
So which paper, guys?

00:02:11.360 --> 00:02:13.120
Mr. Herbert doesn't say.

00:02:13.900 --> 00:02:16.660
And in principle, that's the first lesson of this episode.

00:02:16.900 --> 00:02:20.500
So normally AI synthesis is a footnote-free thing,

00:02:21.140 --> 00:02:24.580
which means that the result it's actually describing comes from

00:02:26.030 --> 00:02:30.260
an interesting paper from 2017 from Fuchi and Nakayama,

00:02:30.640 --> 00:02:34.660
and both are Japanese physicists and both are still very active in the field,

00:02:34.860 --> 00:02:36.740
and they published in Physically Review Applied.

00:02:37.879 --> 00:02:41.180
which is the founding paper of the quantum reservoir computing thing.

00:02:41.920 --> 00:02:46.920
It's by the way a numerical result and not a hardware demonstration at the time.

00:02:47.860 --> 00:02:51.000
And the experimental confirmation came a few years later,

00:02:51.340 --> 00:02:54.440
so the correct claim from the confidence level

00:02:55.600 --> 00:02:55.820
"Missing

00:03:16.880 --> 00:03:21.260
Called a practical few atom quantum reservoir computing was an archive paper published in

00:03:22.220 --> 00:03:31.320
November this year last year you tell right just to make sure you didn't completely misunderstand the math exactly because we're essentially talking about taking the processing power of a skyscraper and

00:03:31.880 --> 00:03:32.200
and

00:03:32.600 --> 00:03:35.380
while finding it natively occurring inside a grain of sand.

00:03:35.780 --> 00:03:39.520
It's mind-bending, and so welcome to this deep dive for Entangle Health.

00:03:42.020 --> 00:03:45.380
May you have this one right? Let's see how it continues.

00:03:45.500 --> 00:03:49.060
If you're joining us today, you're likely a researcher or maybe a clinician,

00:03:50.080 --> 00:03:56.160
bioinformatician, or just someone who's fascinated by where cutting-edge tech actually meets biology.

00:03:56.480 --> 00:03:59.560
Exactly, and our mission today is to really unpack

00:03:59.700 --> 00:04:02.080
this totally counterintuitive concept.

00:04:02.760 --> 00:04:06.980
We want to look at how natural physical systems can inherently compute.

00:04:07.400 --> 00:04:09.000
Which is a huge paradigm shift.

00:04:09.050 --> 00:04:09.959
It really is.

00:04:10.110 --> 00:04:14.640
And specifically, we're going to explore what this bizarre quantum scaling means

00:04:14.950 --> 00:04:20.660
for processing the incredibly chaotic, non-linear biological data we deal with in medicine every day.

00:04:20.820 --> 00:04:23.280
Right, things like brainwaves or heart rhythms.

00:04:23.700 --> 00:04:27.940
Because that biological data is, I mean, let's be honest, it's notoriously difficult to parse.

00:04:28.220 --> 00:04:30.240
Oh, absolutely. It's just noise upon noise.

00:04:31.220 --> 00:04:33.260
Oh, absolutely. It's just noise upon noise.

00:04:34.980 --> 00:04:38.780
Okay. I really want to push back hard on that one from Sheila,

00:04:39.500 --> 00:04:43.240
because if you work with EEG or ECG for a living,

00:04:44.370 --> 00:04:46.520
this will sound almost insulting.

00:04:46.980 --> 00:04:50.400
So brainwaves and heart rhythms are absolutely not just noise.

00:04:50.680 --> 00:04:52.700
Yeah, yes, they are complex, but

00:04:53.120 --> 00:05:00.280
On the other hand, some of the most structured, physiologically meaningful signals, and ECG,

00:05:00.310 --> 00:05:04.340
for example, has a beautiful organized PQRST morphology.

00:05:05.490 --> 00:05:10.020
EEG has really well-defined frequency bands that track slip stages, etc.

00:05:10.980 --> 00:05:16.480
What often looks like noise is usually one of the three things, either it's multiple overlapping

00:05:16.740 --> 00:05:22.240
physiological processes that we haven't separated yet, or it's a technical artifact from movement

00:05:22.630 --> 00:05:26.960
like microphones, muscles and electrodes etc.

00:05:28.190 --> 00:05:31.600
Or it's a model that's just too simple to see the structure.

00:05:32.650 --> 00:05:34.120
So the honest statement isn't.

00:05:34.710 --> 00:05:35.300
Oh, absolutely.

00:05:35.580 --> 00:05:36.880
It's just noise upon noise.

00:05:37.780 --> 00:05:44.780
No, its rather biological signals are non-stationary, multi-scale and heavily superimposed or superposed.

00:05:45.759 --> 00:05:50.020
The baseline drifts as the patient moves or breathes.

00:05:50.590 --> 00:05:57.160
Will them change with circadian phases and drugs and many subsystems are talking at once?

00:05:58.400 --> 00:06:03.080
So from the perspective of a simple linear model, that does look like noise.

00:06:03.840 --> 00:06:08.320
That's exactly where reservoir computing, classical or quantum, becomes interesting

00:06:08.880 --> 00:06:12.620
by mapping those time series into a much higher dimensional dynamical space.

00:06:12.680 --> 00:06:14.419
So let's see if those guys...

00:06:14.400 --> 00:06:18.400
Get it right in the next few minutes or so.

00:06:19.139 --> 00:06:21.960
But for the moment I'd rephrase this completely too.

00:06:22.220 --> 00:06:23.620
It's not that the body is noisy.

00:06:24.100 --> 00:06:29.979
It's rather that our models are often deaf to the structure that is already there.

00:06:30.259 --> 00:06:30.660
So let's see.

00:06:32.120 --> 00:06:37.660
In this story, we've brought these monolithic, classical computer architectures with this sort of brute force mindset.

00:06:38.100 --> 00:06:41.380
expecting that sheer size, more nose, more parameters, more energy,

00:06:41.700 --> 00:06:43.940
is the only path to decoding the human body.

00:06:44.340 --> 00:06:45.980
But quantum mechanics is basically suggesting,

00:06:46.440 --> 00:06:51.380
"Uh, no, there is an entirely different, vastly more efficient blueprint out there."

00:06:51.420 --> 00:06:52.760
Profoundly different blueprints.

00:06:52.840 --> 00:06:54.320
I am so ready to get into this.

00:06:55.120 --> 00:06:56.220
But I think, um...

00:06:56.840 --> 00:07:01.260
To understand how seven mere atoms can rival a massive traditional neural network,

00:07:01.960 --> 00:07:04.480
we can't just jump straight into the quantum realm.

00:07:04.580 --> 00:07:05.800
No, we lose everyone instantly.

00:07:06.320 --> 00:07:10.760
Right. We need a bridge, and that bridge, looking at the sources we've got today,

00:07:11.300 --> 00:07:14.540
is a classical machine learning concept called reservoir computing.

00:07:15.000 --> 00:07:15.120
Yes.

00:07:15.700 --> 00:07:20.860
And I'll be honest, when I was reviewing the material for this, this concept felt like

00:07:20.860 --> 00:07:23.760
a massive departure from everything we hear about AI today.

00:07:25.680 --> 00:07:31.600
Okay, let me pause on that when I was reviewing the material thing, because it's worth being

00:07:31.740 --> 00:07:34.820
transparent about what the material actually is here.

00:07:35.150 --> 00:07:40.240
So the AI hosts Sheila and Himmbert.

00:07:41.340 --> 00:07:42.700
are reading random blog posts.

00:07:43.240 --> 00:07:47.180
So I fed them notebook LM, a very curated bundle,

00:07:47.600 --> 00:07:50.040
my own long form white paper on classical

00:07:50.120 --> 00:07:53.100
and quantum reservoir computing and with the founding,

00:07:53.500 --> 00:07:56.540
Fujinakayama paper, next generation reservoir computing

00:07:56.740 --> 00:08:00.280
work from Gauti and colleagues and several review articles

00:08:00.500 --> 00:08:00.700
and,

00:08:01.000 --> 00:08:03.600
physical and neuromorphic reservoir computing.

00:08:04.200 --> 00:08:07.880
Oh, it's basically a rebellion against the modern deep learning paradigm.

00:08:08.420 --> 00:08:12.840
I mean, to understand why it's so radical, you kind of have to look at what everybody else is doing right now.

00:08:13.660 --> 00:08:17.600
And what's interesting to me as the person who composed the source material,

00:08:18.660 --> 00:08:25.600
which was used to the train this notebook LM, is that this thing actually did pick up the right

00:08:25.740 --> 00:08:27.600
intellectual contrast. So brute force.

00:08:28.100 --> 00:08:32.620
fully trained networks versus let's the physics do the work and only read out the edges.

00:08:32.849 --> 00:08:33.479
So let's see.

00:08:34.919 --> 00:08:38.400
The dominant approach to sequential data.

00:08:38.750 --> 00:08:43.020
Like say, reading a sentence or tracking a patient's vital signs over a week.

00:08:43.300 --> 00:08:43.539
Exactly.

00:08:43.970 --> 00:08:48.960
For that kind of time series data, everyone uses recurrent neural networks or more recently

00:08:49.110 --> 00:08:50.400
these massive transformers.

00:08:50.860 --> 00:08:59.020
the architectures behind all the famous chatbots right all the predictive models making headline yeah exactly those and in those standard architectures you have this incredibly.

00:09:00.080 --> 00:09:03.700
Dense network of artificial neurons and to teach that network.

00:09:04.320 --> 00:09:15.100
How to actually do a task you have to painfully pin and staking optimize every single internal connection between those neurons which takes forever right when the network makes a mistake.

00:09:15.500 --> 00:09:19.080
There's this mathematical algorithm called backpropagation.

00:09:19.580 --> 00:09:24.500
It runs and revolves through the entire network and it tweaks millions or even billions of

00:09:24.560 --> 00:09:27.560
microscopic weights just to minimize the error for the next time.

00:09:30.220 --> 00:09:33.000
Maybe a quick meta comment here generally.

00:09:33.400 --> 00:09:37.420
So we are in 10 minutes into the series and if you're regular entangled health listener

00:09:38.180 --> 00:09:50.279
You've probably noticed that this feels quite different from the usual episodes and it's much more like a structured educational lesson than the normal freeform philosophy conversation and this is not an accident.

00:09:51.579 --> 00:09:59.380
I really deliberately asked notebook alarm to build the dialectic step-by-step tutorial on reservoir computing and quantum reservoir computing.

00:10:00.420 --> 00:10:06.140
And you can hear that in the pacing and the amount of hand holding, analogies, etc.

00:10:07.020 --> 00:10:11.380
I think it's a bit overdone, but let's see where this brings us.

00:10:11.560 --> 00:10:13.340
We have some time in front of us.

00:10:14.840 --> 00:10:19.020
So it's really a different flavour of entangled health, more classroom than coffee, and the

00:10:19.480 --> 00:10:20.660
philosophical layer comes eventually.

00:10:21.100 --> 00:10:25.320
Okay, so if I can try an analogy here, it feels a bit like trying to coordinate a stadium

00:10:25.320 --> 00:10:28.840
wave by giving a highly specific individualized script to

00:10:31.120 --> 00:10:38.100
There's also fits, but the philosophical layer comes in through what we are doing right now.

00:10:38.380 --> 00:10:44.740
There's me stopping the tape actually, challenging the framing and asking what happens when I

00:10:44.839 --> 00:10:47.140
feel that I teach its own foundations.

00:10:49.140 --> 00:10:51.780
We have to re-hear this, but the quote was quite fun.

00:10:52.220 --> 00:10:56.920
And it tweaks millions, or even billions, of microscopic weights, just to minimize the

00:10:57.000 --> 00:10:57.840
error for the next time.

00:10:58.180 --> 00:11:02.200
Okay, so if I can try an analogy here, it feels a bit like trying to coordinate a stadium

00:11:02.580 --> 00:11:05.960
wave by giving a highly specific individualized script to

00:11:06.400 --> 00:11:08.340
I don't know, 80,000 people in the seats.

00:11:08.720 --> 00:11:10.540
That is a brilliant way to picture it, actually.

00:11:10.720 --> 00:11:12.780
And if the wave looks a bit sloppy, you have to stop everything,

00:11:13.200 --> 00:11:18.140
go back and tell person number 4,302 to stand up a millisecond later,

00:11:18.540 --> 00:11:21.900
and tell person 12,050 to lift their arms slightly higher.

00:11:22.240 --> 00:11:24.840
Yes. And you have to do that over and over again.

00:11:25.000 --> 00:11:26.540
It requires massive amounts of data,

00:11:27.060 --> 00:11:28.740
staggering amounts of computational time,

00:11:29.240 --> 00:11:31.100
and likely said huge energy consumption.

00:11:32.260 --> 00:11:35.140
Yeah, this was the classical machine learning part, just to remind you.

00:11:36.260 --> 00:11:37.300
Huge energy consumption.

00:11:37.480 --> 00:11:38.520
It sounds exhausting.

00:11:38.620 --> 00:11:38.980
It is.

00:11:39.740 --> 00:11:44.160
But back in 2001 and 2002, there were these two researchers, Herbert Yeager and Wolfgang

00:11:44.440 --> 00:11:44.580
Mass.

00:11:45.280 --> 00:11:50.600
They independently looked at this whole messy process and basically thought, "Um,

00:11:51.000 --> 00:11:53.860
What if we just don't bother training the internal network at all?

00:11:54.660 --> 00:11:56.040
Which sounds completely absurd.

00:11:56.260 --> 00:11:59.820
I mean, if you don't train the AI, how on earth does it learn anything?

00:12:00.080 --> 00:12:02.580
Well, this is the sheer elegance of reservoir computing.

00:12:03.220 --> 00:12:04.960
The architecture basically just has three parts.

00:12:05.220 --> 00:12:08.940
You've got an input layer, a hidden layer, which is what we call the reservoir, and an

00:12:09.040 --> 00:12:09.760
output readout layer.

00:12:10.040 --> 00:12:10.980
Okay, simple enough.

00:12:11.220 --> 00:12:11.300
Right.

00:12:11.960 --> 00:12:15.220
And Yeager, he called his version "Echo State Networks."

00:12:15.420 --> 00:12:17.120
Use these artificial math-based nodes.

00:12:18.220 --> 00:12:22.840
Meanwhile, Mass called his liquid state machines because he was drawing inspiration directly

00:12:23.080 --> 00:12:24.180
from biological brain neurons.

00:12:24.930 --> 00:12:26.740
But the underlying idea is the same.

00:12:27.030 --> 00:12:29.040
The core principle is completely identical.

00:12:29.890 --> 00:12:33.480
You generate the internal connections of your reservoir completely randomly.

00:12:34.100 --> 00:12:36.139
And this is the crazy part.

00:12:36.660 --> 00:12:38.279
You lock them, you never tweak them,

00:12:38.620 --> 00:12:39.319
they're completely fixed.

00:12:39.620 --> 00:12:41.699
- Oh wait, you just wire a bunch of nodes together

00:12:41.800 --> 00:12:43.319
at random, throw your beta in,

00:12:43.920 --> 00:12:45.100
but just hope for the best.

00:12:45.440 --> 00:12:46.620
- Well you don't hope for the best.

00:12:47.139 --> 00:12:48.519
You rely on the natural physics

00:12:48.740 --> 00:12:50.199
of a complex dynamical system.

00:12:50.399 --> 00:12:51.500
- Okay, unpack that for me.

00:12:54.940 --> 00:12:56.020
Okay, unpack that for me.

00:12:56.280 --> 00:12:59.540
So you feed your input, let's say it's a patient's continuous heart rate monitor data

00:12:59.760 --> 00:13:01.420
into this fixed random reservoir.

00:13:02.830 --> 00:13:04.980
The reservoir's job is not to give you the answer.

00:13:05.280 --> 00:13:06.780
It's not, then what is it doing?

00:13:07.270 --> 00:13:12.480
The reservoir's job is simply to map that low dimensional input into a massive, highly

00:13:12.620 --> 00:13:14.220
complex, high dimensional space.

00:13:14.800 --> 00:13:20.440
The only part of the entire system that you actually train is the final linear readout layer.

00:13:20.760 --> 00:13:24.980
Okay, let's slow down and translate that a bit because mapping to a high dimensional space

00:13:25.180 --> 00:13:29.180
is exactly the kind of jargon that makes people tune out.

00:13:29.380 --> 00:13:30.000
Fair enough, yeah.

00:13:30.240 --> 00:13:34.720
And this brings us to the stone and upon analogy, which is great because

00:13:35.100 --> 00:13:38.440
It was literally the original inspiration for Wolfgang Mass.

00:13:38.740 --> 00:13:41.500
It really is the perfect physical representation of the math.

00:13:41.899 --> 00:13:45.339
So imagine you drop a stone into a perfectly still pond.

00:13:45.940 --> 00:13:47.920
The stone represents your input data.

00:13:48.160 --> 00:13:51.839
Let's say it's a specific spike on an ECG, an electrocardiogram.

00:13:52.139 --> 00:13:52.899
Right, the initial pulse.

00:13:53.360 --> 00:13:57.240
When that stone hits the water, it doesn't just make one tiny splash.

00:13:57.920 --> 00:14:02.040
It creates this complex, cascading, interacting pattern of ripples.

00:14:02.500 --> 00:14:04.820
that spreads across the entire surface.

00:14:05.480 --> 00:14:07.660
So the pond itself is our fixed reservoir.

00:14:08.100 --> 00:14:10.100
-Yes. And notice what the pond is actually doing there.

00:14:10.220 --> 00:14:12.580
It's taking a very simple single piece of information,

00:14:12.880 --> 00:14:15.320
the impact of the stone, and transforming it

00:14:15.420 --> 00:14:18.520
into an incredibly rich, complex, physical pattern

00:14:19.000 --> 00:14:19.900
across a wide area.

00:14:20.200 --> 00:14:21.880
-Automatically, you don't have to train the water.

00:14:22.100 --> 00:14:25.200
-Exactly. The physics of the water handles the complexity for free.

00:14:25.440 --> 00:14:27.760
-Right. And so if I'm a doctor and I'm trying to figure out

00:14:27.920 --> 00:14:30.200
how big that stone was or how fast it was moving

00:14:30.360 --> 00:14:30.860
when I hit the water,

00:14:31.360 --> 00:14:37.040
I do not need a supercomputer to calculate the individual trajectory of every single water molecule in the pond.

00:14:37.560 --> 00:14:38.760
No, because that would be back propagation.

00:14:39.380 --> 00:14:39.620
Exactly.

00:14:40.500 --> 00:14:43.480
Instead, I just place a few simple sensors at the very edge of the pond.

00:14:43.640 --> 00:14:44.420
That's my readout layer.

00:14:44.820 --> 00:14:47.840
I just take a snapshot of the ripple patterns as they wash up against the shore.

00:14:48.620 --> 00:14:52.740
And by training only those edge sensors to recognize specific ripple patterns,

00:14:53.250 --> 00:14:55.800
you can deduce everything about the original stone.

00:14:56.339 --> 00:14:58.300
You just use a very simple mathematical tool,

00:14:58.920 --> 00:15:03.380
usually a linear regression, or what's known in the literature as a pseudo-inverse calculation.

00:15:04.100 --> 00:15:08.459
Okay, let's quickly define pseudo-inverse for the clinician listening who maybe hasn't

00:15:08.540 --> 00:15:10.880
taken linear algebra in 20 years.

00:15:10.959 --> 00:15:11.519
Oh sure, yeah.

00:15:12.320 --> 00:15:17.180
Essentially, a pseudo-inverse calculation is just finding the line of best fit through

00:15:17.180 --> 00:15:21.940
a messy scatter plot of data, a very cheap, very fast mathematical operation.

00:15:22.360 --> 00:15:25.519
So you're basically just assigning basic weights to those final edge sensors.

00:15:26.019 --> 00:15:26.339
That's it.

00:15:26.600 --> 00:15:29.820
You get the immense complex processing power of the entire pond.

00:15:30.300 --> 00:15:35.580
But the only computational cost you actually pay is the energy it takes to read the sensors at the edge.

00:15:36.440 --> 00:15:39.260
The pond does all the heavy lifting naturally just by being water.

00:15:40.500 --> 00:15:41.680
Okay, let's listen to this one.

00:15:42.399 --> 00:15:45.300
The pond does all the heavy lifting naturally just by being water.

00:15:45.820 --> 00:15:48.560
Yes, but it's not so easy to implement this as a device.

00:15:49.140 --> 00:15:50.100
Okay, I love this.

00:15:50.899 --> 00:15:54.680
But the pond has to behave in a very specific way for this to work, right?

00:15:55.579 --> 00:15:59.620
The literature keeps bringing up this term, the "Echo-State" property, or fading memory.

00:15:59.980 --> 00:16:01.220
Yes, that is crucial.

00:16:01.899 --> 00:16:05.360
Because, I mean, if I throw a stone in and the ripples bounce back and forth off the

00:16:05.459 --> 00:16:09.600
shores forever without stopping, what happens when I throw the next stone in a second later?

00:16:09.899 --> 00:16:11.000
You get a chaotic mess.

00:16:11.240 --> 00:16:11.459
Right.

00:16:11.880 --> 00:16:16.540
If the water never settles down, the new ripples collide with the infinite old ripples.

00:16:17.970 --> 00:16:22.380
Very quickly, the surface of the pond becomes pure, unpredictable turbulence.

00:16:23.519 --> 00:16:28.120
If you try to read the sensors at the edge, you have absolutely no idea if you're looking

00:16:28.180 --> 00:16:31.959
at the echo of stone number one or stone number two or stone number 500.

00:16:32.579 --> 00:16:35.139
So the system basically becomes too chaotic to be useful.

00:16:35.740 --> 00:16:35.839
Exactly.

00:16:35.940 --> 00:16:37.620
So the pond must eventually calm down.

00:16:38.160 --> 00:16:39.720
The memory of the stone has to fade.

00:16:40.139 --> 00:16:40.519
Feeding memory.

00:16:40.759 --> 00:16:41.420
Okay, I get it.

00:16:41.699 --> 00:16:43.199
But it can't fade too quickly either, right?

00:16:43.240 --> 00:16:43.920
No, it can't.

00:16:44.380 --> 00:16:53.079
Imagine if you replace the water in the pond with

00:16:46.400 --> 00:16:48.060
like thick, cold molasses.

00:16:48.340 --> 00:16:49.100
Oh, wow. Okay.

00:16:49.170 --> 00:16:50.660
You drop the stone in and it just goes

00:16:51.080 --> 00:16:53.300
for a drink and no ripple sprayed at all.

00:16:53.580 --> 00:16:53.680
Exactly.

00:16:54.800 --> 00:16:56.440
The sensors at the edge detect nothing.

00:16:57.360 --> 00:17:00.800
In that scenario, the system has no short-term memory at all.

00:17:01.560 --> 00:17:05.420
It can't process sequential data because by the time the second input arrives,

00:17:05.800 --> 00:17:09.140
the system has completely forgotten that the first input ever existed.

00:17:09.720 --> 00:17:13.100
So the perfect reservoir has to exist in a kind of goldilocks zone,

00:17:13.600 --> 00:17:15.820
not too chaotic, not too rigid.

00:17:16.180 --> 00:17:20.120
Precisely. In Dynamical Systems Theory, we actually call this the edge of chaos.

00:17:20.520 --> 00:17:22.140
Edge of chaos, that's a great term.

00:17:22.500 --> 00:17:22.860
It is.

00:17:23.600 --> 00:17:27.620
The reservoir needs just enough memory to retain information about recent inputs.

00:17:28.080 --> 00:17:31.860
It only echoes the past, so it can contextualize what's happening right now.

00:17:32.440 --> 00:17:34.640
But it must gradually forget older inputs.

00:17:35.100 --> 00:17:38.100
So the state of the pond is uniquely determined by recent history.

00:17:38.520 --> 00:17:40.240
Not by something that happens, say...

00:17:40.460 --> 00:17:43.820
an hour ago. Okay, I feel like we have a very solid grasp on the classical pond now.

00:17:44.220 --> 00:17:48.680
Whether that pond is made of actual water or a random network of artificial neurons or even a

00:17:48.720 --> 00:17:53.360
network of optical fibers, the core principle completely holds. It does. It's universally applicable

00:17:53.740 --> 00:17:58.220
to dynamical systems. But this deep dive is for entangled health. We didn't come here just to talk

00:17:58.360 --> 00:18:04.220
about classical physics. No, definitely not. So we have to ask the critical question. What happens when

00:18:04.220 --> 00:18:07.700
when we take that classical pond and we swap out the water.

00:18:08.100 --> 00:18:12.740
for a medium that is governed by the bizarre probabilistic laws of quantum mechanics.

00:18:14.780 --> 00:18:18.060
Well, then we enter the realm of quantum reservoir computing, or QRC,

00:18:18.780 --> 00:18:21.740
and this is exactly how we answer that hook from the beginning of the show.

00:18:22.240 --> 00:18:25.080
The 5 to 7 qubits rivaling 500 nodes.

00:18:25.520 --> 00:18:30.680
Exactly. It all comes down to the fundamental difference between classical bits and quantum states.

00:18:31.320 --> 00:18:33.240
It's a concept known as Hilbert space scaling.

00:18:33.580 --> 00:18:34.920
Okay, let's unpack Hilbert space.

00:18:35.200 --> 00:18:38.860
Because that is a heavy, intimidating physics term.

00:18:39.529 --> 00:18:41.220
Break it down for me like I'm new to this.

00:18:41.500 --> 00:18:44.980
Sure. Let's stick with the idea of a physical space where our data can exist.

00:18:45.620 --> 00:18:48.940
In a classical system, let's say our artificial neural network,

00:18:49.659 --> 00:18:52.740
if you have 10 nodes, your state-space dimension is 10.

00:18:53.019 --> 00:18:54.340
You have 10 variables you can play with.

00:18:54.480 --> 00:18:55.679
Right. It scales linearly.

00:18:56.100 --> 00:18:58.799
If you want more complexity, you have to physically build more nodes.

00:18:59.000 --> 00:19:01.660
Ten nodes, ten dimensions, 100 nodes, 100 dimensions.

00:19:02.000 --> 00:19:02.900
It's very straightforward.

00:19:03.500 --> 00:19:06.100
But a quantum system does not scale linearly.

00:19:06.830 --> 00:19:11.059
A quantum system consisting of n quibits spans a Hilbert space

00:19:11.500 --> 00:19:15.240
which is just a mathematical space describing all its possible states of dimension

00:19:15.940 --> 00:19:17.120
2 to the power of n.

00:19:17.520 --> 00:19:19.440
2 to the power of n, wait, so it's exponential?

00:19:19.900 --> 00:19:21.300
It is entirely exponential.

00:19:21.680 --> 00:19:25.780
If you have just 10 qubits, your dimensional space isn't 10, it's 1024.

00:19:26.700 --> 00:19:27.040
Wow.

00:19:27.620 --> 00:19:31.600
Yeah, and if you have 50 qubits, your dimension is 2 to the power of 50.

00:19:32.000 --> 00:19:38.740
That is a number so large that no classical supercomputer currently on Earth has enough memory to even simulate it

00:19:38.980 --> 00:19:46.660
That's insane. It is. The quantum physical system inherently provides this exponentially massive feature space basically for free

00:19:47.660 --> 00:19:49.900
That is real and it's really impressive

00:19:50.060 --> 00:19:58.120
But the spoiler the Hilbert space being that large and actually being able to use it are two very different things more than that in a second

00:19:59.000 --> 00:20:05.140
The quantum physical system inherently provides this exponentially massive feature space, basically for free.

00:20:05.980 --> 00:20:07.600
Simply by existing as a quantum object.

00:20:07.880 --> 00:20:13.960
Okay, I really want to make sure I'm visualizing this correctly because in a lot of pop science, we hear quantum entanglement described as this.

00:20:14.860 --> 00:20:17.200
sort of telepathic whispering network,

00:20:17.440 --> 00:20:19.260
like the atoms are magically talking to each other

00:20:19.400 --> 00:20:20.460
instantly across space,

00:20:20.910 --> 00:20:22.660
but that's not really how a computer says it.

00:20:22.760 --> 00:20:23.940
- I am so glad you brought that up

00:20:24.040 --> 00:20:25.600
because that magic telepathy trope

00:20:25.930 --> 00:20:27.500
is one of the most frustrating misconceptions

00:20:27.890 --> 00:20:28.540
in quantum science.

00:20:28.940 --> 00:20:29.560
- It's everywhere.

00:20:30.060 --> 00:20:30.620
- It really is.

00:20:31.150 --> 00:20:32.060
But it's not magic.

00:20:32.660 --> 00:20:34.940
It's about wave interference and shared probabilities.

00:20:35.540 --> 00:20:37.900
- So how should we picture the quantum pond then?

00:20:38.300 --> 00:20:41.620
- Okay, imagine instead of dropping a stone into water.

00:20:42.320 --> 00:20:44.580
you're dropping an acoustic tone, a sound wave,

00:20:45.020 --> 00:20:47.440
into a room filled with perfectly tuned tuning forks.

00:20:47.760 --> 00:20:48.500
Tuning forks, okay.

00:20:48.600 --> 00:20:51.220
In classical system, each tuning fork only vibrates

00:20:51.300 --> 00:20:54.600
if you hit its exact specific frequency, right?

00:20:55.680 --> 00:20:57.000
And they don't really interact with each other.

00:20:57.100 --> 00:20:57.540
Right, they're isolated.

00:20:58.160 --> 00:21:00.640
But in a quantum system, because of entanglement,

00:21:01.160 --> 00:21:03.920
all the tuning forks are physically coupled in their probabilities.

00:21:04.600 --> 00:21:08.380
So when the input tone enters the room, it doesn't just make one fork ring,

00:21:09.240 --> 00:21:11.559
it creates a massive instantaneous chord.

00:21:12.000 --> 00:21:12.500
Oh, interesting.

00:21:12.960 --> 00:21:15.919
Yeah, the phase of one wave directly interferes with the phase of another.

00:21:16.280 --> 00:21:20.340
They amplify each other, they cancel each other out, and they create these incredibly complex,

00:21:20.940 --> 00:21:23.140
multilayered harmonies completely automatically.

00:21:24.000 --> 00:21:27.960
So the nonlinear transformation, the complex ripples we needed for our reservoir,

00:21:28.700 --> 00:21:31.600
happens simply through the physical physics of quantum wave interference.

00:21:32.320 --> 00:21:37.240
Exactly. You don't need millions of artificial neurons calculating math. You just let the quantum

00:21:37.380 --> 00:21:43.300
waves crash into each other. The quibits basically evolve under what physicists call a Hamiltonian.

00:21:43.500 --> 00:21:48.880
And a Hamiltonian is? It's just a mathematical description of the system's total energy and how

00:21:48.880 --> 00:21:51.300
the particles interact over time. This Hamiltonian,

00:21:51.600 --> 00:21:55.860
naturally mixes the input information across that exponentially large,

00:21:56.360 --> 00:21:58.039
head and 24-dimensional Hilbert space.

00:21:58.500 --> 00:22:00.520
Okay, but hold on, let me push back on this a bit,

00:22:00.640 --> 00:22:05.100
because I know exactly how dangerous exponential hype can be in tech.

00:22:05.299 --> 00:22:05.840
Oh, absolutely.

00:22:06.320 --> 00:22:08.140
You're saying the space scales exponentially.

00:22:08.799 --> 00:22:12.740
Does that mean if we just keep adding quibits and making the Hamiltonian more and more complex,

00:22:13.100 --> 00:22:15.679
the reservoir automatically becomes infinitely smarter.

00:22:16.520 --> 00:22:18.419
Like, if we build a hundred kweeb at pond,

00:22:18.980 --> 00:22:24.140
does it suddenly have the expressive power to solve literally any biological data problem instantly?

00:22:24.520 --> 00:22:25.880
I mean, what is stopping us?

00:22:26.020 --> 00:22:28.200
That is a phenomenal rigorous question.

00:22:28.500 --> 00:22:32.100
And to answer it, we have to look closely at a critical finding from the literature.

00:22:33.290 --> 00:22:38.180
Specifically, there's a recent paper we reviewed titled, "Expressive Limits of Quantum Reservoir Computing."

00:22:38.420 --> 00:22:39.360
Right, I remember that one.

00:22:39.680 --> 00:22:41.900
And the short answer to your question is, "No."

00:22:42.060 --> 00:22:46.000
No. So an exponentially larger space does not equal exponential intelligence.

00:22:46.460 --> 00:22:47.460
It absolutely does not.

00:22:48.500 --> 00:22:49.680
And the short answer to your question is...

00:22:50.000 --> 00:22:54.900
No. No. So an exponentially larger space is not equal exponential intelligence.

00:22:55.300 --> 00:22:56.320
It absolutely does not.

00:22:57.440 --> 00:23:00.200
And this is a crucial reality check for the entire field.

00:23:01.160 --> 00:23:07.280
We have to discuss why more reservoir complexity does not automatically increase the expressivity of the quantum system.

00:23:07.480 --> 00:23:08.100
Okay, why not?

00:23:08.500 --> 00:23:16.900
The literature mathematically proves that the expressive power of a quantum reservoir is strictly limited by how the input is injected into the system in the first place.

00:23:17.100 --> 00:23:20.300
Ah, by how we drop the stone into the pond.

00:23:20.700 --> 00:23:21.280
Precisely.

00:23:22.020 --> 00:23:24.740
Think about how data actually enters a quantum computer.

00:23:25.740 --> 00:23:30.760
You usually encode classical data like our heart rate monitor numbers by applying single-quivot

00:23:30.880 --> 00:23:31.260
rotations.

00:23:31.690 --> 00:23:36.160
You shoot a very specific microwave pulse at a single atom to change its spin.

00:23:36.930 --> 00:23:41.740
The paper demonstrates that the unitary dynamics of the reservoir, meaning the natural closed-loop

00:23:41.940 --> 00:23:44.420
way the quantum states evolve and interact over time,

00:23:44.900 --> 00:23:46.920
only mix those inputs linearly.

00:23:47.240 --> 00:23:47.940
Wait, wait, I'm confused.

00:23:48.070 --> 00:23:53.240
I thought the whole point of the quantum pond was that it gives us massively complex nonlinear processing.

00:23:53.860 --> 00:23:55.460
Well, it provides a massive space, yes.

00:23:56.560 --> 00:24:01.140
But we have to be highly specific about where the nonlinearity actually comes from.

00:24:01.280 --> 00:24:01.500
Okay.

00:24:02.120 --> 00:24:06.900
The Reservoir's internal evolution mixes the features in a vastly high dimensional space.

00:24:08.000 --> 00:24:13.100
But it cannot magically generate higher order nonlinearities if the initial input encoding

00:24:13.500 --> 00:24:14.200
doesn't support it.

00:24:14.360 --> 00:24:15.040
Oh, I see.

00:24:15.550 --> 00:24:21.380
The expressivity, meaning the maximum number of distinct orthogonal mathematical functions

00:24:21.430 --> 00:24:22.460
the system can express.

00:24:22.900 --> 00:24:27.240
has a hard upper limit, and that ceiling is dictated entirely by the input layer.

00:24:27.640 --> 00:24:31.720
So if you feed the quantum system's simple data in a very simple linear way,

00:24:31.840 --> 00:24:34.460
it doesn't matter how beautifully entangled the waves are inside.

00:24:34.820 --> 00:24:37.980
The system can only ever output variations of that simple data.

00:24:38.220 --> 00:24:38.380
Exactly.

00:24:38.580 --> 00:24:40.680
The bottleneck is literally the front door.

00:24:41.000 --> 00:24:41.900
That's a great way to put it.

00:24:42.200 --> 00:24:45.940
Increasing the complexity of the reservoir's internal Hamiltonian beyond a certain point

00:24:46.080 --> 00:24:47.020
does absolutely nothing.

00:24:47.300 --> 00:24:48.840
to increase its expressivity.

00:24:49.010 --> 00:24:50.760
In fact, it might actually be detrimental.

00:24:51.040 --> 00:24:51.280
- Big instrumental.

00:24:51.720 --> 00:24:53.360
- Because it just scrambles the information

00:24:53.560 --> 00:24:54.680
into unreadable noise.

00:24:55.700 --> 00:24:58.040
To truly unlock the power of QRC,

00:24:58.800 --> 00:25:01.580
researchers are finding they have to pre-process the data.

00:25:02.310 --> 00:25:04.520
They need non-linear input transformations

00:25:05.080 --> 00:25:08.260
or they need to map the inputs to higher order observables.

00:25:08.820 --> 00:25:11.720
You cannot just treat the quantum chip as a magical black box

00:25:11.940 --> 00:25:14.000
that will figure out the complexity for you.

00:25:14.200 --> 00:25:14.480
Exactly.

00:25:14.920 --> 00:25:19.080
You have to design the input stage with just as much care as the reservoir itself.

00:25:19.300 --> 00:25:21.480
And that is such an important grounding point.

00:25:21.780 --> 00:25:23.240
The magic is an infinite.

00:25:23.700 --> 00:25:26.820
But, okay, let's assume we are brilliant engineers.

00:25:27.419 --> 00:25:32.120
We have optimized our input encoding, we drop the stone perfectly, and the quantum ripples

00:25:32.659 --> 00:25:36.720
spread through this massive Hilbert space, generating exactly the complex features we

00:25:36.779 --> 00:25:36.919
need.

00:25:37.100 --> 00:25:38.620
Okay, so we have a perfect quantum pond.

00:25:38.940 --> 00:25:42.580
Right, now we hit what has to be the most famous roadblock in all of quantum mechanics.

00:25:43.320 --> 00:25:46.740
Because to use the reservoir, we actually have to read the sensors at the edge.

00:25:46.860 --> 00:25:47.899
We have to measure the ripples.

00:25:48.159 --> 00:25:49.580
Ah, yes, the measurement problem.

00:25:50.100 --> 00:25:50.880
The observer effect.

00:25:51.300 --> 00:25:52.920
Shortening your absolute nightmare.

00:25:53.300 --> 00:25:53.540
Exactly.

00:25:54.280 --> 00:25:58.600
Standard quantum mechanics tells us that the moment you observe a quantum system, you force

00:25:58.700 --> 00:26:00.160
it to pick a definitive state.

00:26:00.800 --> 00:26:01.820
The way function collapses.

00:26:02.740 --> 00:26:07.180
So if you take a snapshot of the quantum pawn, you instantly freeze the entire thing

00:26:07.200 --> 00:26:08.000
into solid ice.

00:26:08.200 --> 00:26:09.880
you destroy the quantum coherence,

00:26:10.440 --> 00:26:11.960
and critically for our reservoir,

00:26:12.679 --> 00:26:15.000
you completely destroy the fading memory

00:26:15.179 --> 00:26:16.720
we spent so long talking about.

00:26:17.019 --> 00:26:19.659
- It is the single biggest existential threat

00:26:19.779 --> 00:26:20.659
to quantum reservoir computing.

00:26:21.539 --> 00:26:22.700
I mean, think about it mechanically.

00:26:23.660 --> 00:26:27.960
Fading memory relies on the system evolving continuously over time.

00:26:28.680 --> 00:26:32.900
The ripples from heartbeat one need to still be physically moving when heartbeat two arrives.

00:26:33.760 --> 00:26:38.180
But if every time you measure the system to get your readout at, say, time step T,

00:26:38.720 --> 00:26:40.980
you collapse it back to a rigid classical state,

00:26:41.760 --> 00:26:42.900
Well, it has absolute amnesia.

00:26:43.440 --> 00:26:46.720
It has no memory of what happened at T-1 when you feed it the next input.

00:26:47.020 --> 00:26:49.480
It's like resetting the entire universe every single time you blink.

00:26:49.860 --> 00:26:51.880
How on earth do researchers actually solve this?

00:26:52.020 --> 00:26:55.820
Because obviously from the sources we read, they are building these things and making them work.

00:26:56.100 --> 00:26:56.400
They are.

00:26:56.820 --> 00:26:59.320
And the engineering workarounds are honestly incredible.

00:27:00.340 --> 00:27:02.760
Let's break down the three main protocols detailed in the literature.

00:27:03.100 --> 00:27:03.280
Okay.

00:27:03.520 --> 00:27:08.400
The first one is the most conceptually simple, but it's basically a brute force approach.

00:27:09.000 --> 00:27:11.720
It's called the restarting protocol or RSP.

00:27:12.200 --> 00:27:14.200
I can already tell from the name that I'm going to hate this one.

00:27:14.360 --> 00:27:15.680
Oh, it is computationally exhausting.

00:27:16.360 --> 00:27:21.940
In RSP, to process a time series, you start at the very beginning of the patient's data.

00:27:22.740 --> 00:27:26.940
You run the quantum evolution up to point A, you measure it, record the classical result,

00:27:27.490 --> 00:27:28.140
and the system collapses.

00:27:28.720 --> 00:27:29.420
Pawn turns to ice.

00:27:29.900 --> 00:27:30.600
Pawn turns to ice.

00:27:31.440 --> 00:27:35.280
Now to find out what happens at point B, you can't just unfreeze it.

00:27:35.510 --> 00:27:38.520
You have to completely reset the physical quantum system back to zero.

00:27:39.000 --> 00:27:43.980
Then, you feed in all the inputs from the beginning up to point A again, then feed in

00:27:44.040 --> 00:27:47.720
the new input for point B, let it evolve, measure, and collapse again.

00:27:47.940 --> 00:27:48.260
Wait, what?

00:27:48.440 --> 00:27:52.220
You are literally re-evolving the system from scratch for every single tiny data point.

00:27:52.280 --> 00:27:52.820
Every single one.

00:27:53.310 --> 00:27:57.400
That completely defeats the purpose of having an efficient, continuous processor.

00:27:58.659 --> 00:28:03.080
If I'm in an ICU, monitoring a patient's continuous blood pressure for 24 hours,

00:28:03.559 --> 00:28:09.139
I cannot restart the entire mathematical simulation from second zero every single time a new heartbeat happens.

00:28:09.399 --> 00:28:10.059
No, you can't.

00:28:10.059 --> 00:28:12.139
The machine would lag behind reality instantly.

00:28:12.580 --> 00:28:14.960
Exactly. The time complexity scales horribly.

00:28:15.460 --> 00:28:19.360
It is purely a laboratory proof of concept. It just proves the math works.

00:28:19.800 --> 00:28:21.600
So researchers had to get smarter.

00:28:22.020 --> 00:28:22.500
Thank goodness.

00:28:22.980 --> 00:28:25.580
They developed what's called the online protocol, or OLP.

00:28:26.340 --> 00:28:31.760
And this protocol utilizes a fascinating concept called weak measurements, or indirect measurements.

00:28:32.080 --> 00:28:37.980
Okay, I need an ELI-5, explain it like I'm 5 for weak measurements, because if observing the pond freezes it into ice,

00:28:38.300 --> 00:28:39.940
How do you weekly observe it?

00:28:39.980 --> 00:28:41.360
You just squint at it really fast.

00:28:41.600 --> 00:28:44.740
You know, squinting at it is actually a remarkably accurate way to think about it.

00:28:44.880 --> 00:28:45.000
Really?

00:28:45.420 --> 00:28:45.540
Yeah.

00:28:46.240 --> 00:28:47.560
Mechanically, here's what you do.

00:28:47.960 --> 00:28:52.180
Instead of measuring your primary reservoir qubits directly, you introduce a second set

00:28:52.180 --> 00:28:54.460
of qubits called ancillary or helper qubits.

00:28:54.580 --> 00:28:55.600
Helper qubits, okay.

00:28:55.900 --> 00:29:01.100
You gently couple the helper qubits to the reservoir qubits, so they interact just a

00:29:01.140 --> 00:29:01.380
little bit.

00:29:01.800 --> 00:29:04.140
Then, you only measure the helpers.

00:29:04.740 --> 00:29:05.340
Oh, wow.

00:29:05.620 --> 00:29:09.100
Okay, so you're looking at the shadow of the pond rather than the pond itself.

00:29:09.440 --> 00:29:11.100
Yes, beautifully put.

00:29:11.750 --> 00:29:15.400
By measuring the helper qubits, you extract a small amount of classical information about

00:29:15.410 --> 00:29:16.100
the state of the reservoir.

00:29:17.270 --> 00:29:22.400
But because you aren't measuring the reservoir directly, you don't cause a total wave function

00:29:22.580 --> 00:29:25.180
collapse, you cause a partial collapse.

00:29:25.220 --> 00:29:25.960
A partial collapse.

00:29:26.260 --> 00:29:26.400
Right.

00:29:26.780 --> 00:29:29.000
It is an incredibly delicate balancing act.

00:29:29.300 --> 00:29:33.380
The engineers have to perfectly balance the amount of information they extract with the

00:29:33.440 --> 00:29:38.500
degree of disturbance they cause, preserving just enough quantum coherence so the ripples

00:29:38.660 --> 00:29:40.820
keep moving, maintaining that fading memory.

00:29:41.140 --> 00:29:44.480
So you get a blurry picture of the ripples, but the water stays liquid.

00:29:44.980 --> 00:29:45.520
That is ingenious.

00:29:45.960 --> 00:29:46.200
It is.

00:29:47.600 --> 00:29:51.419
It's not perfect. Even weak measurements introduce some back action.

00:29:51.639 --> 00:29:51.760
Back?

00:29:52.120 --> 00:29:55.580
Yeah, it's a degree of quantum decoherence that acts like friction on the water.

00:29:56.159 --> 00:29:57.919
It degrades the memory over time.

00:29:58.139 --> 00:30:00.679
So to truly protect the system from the observer effect,

00:30:01.139 --> 00:30:04.340
the absolute cutting edge of the field is moving toward a third approach.

00:30:05.100 --> 00:30:07.000
It's called feedback-driven QRC.

00:30:07.640 --> 00:30:08.180
Feedback loops.

00:30:08.420 --> 00:30:10.660
Okay, I understand those in classical biology,

00:30:10.800 --> 00:30:12.880
like how the body regulates its own temperature.

00:30:13.540 --> 00:30:17.220
But how does feedback work to stop quantum collapse?

00:30:17.720 --> 00:30:20.740
One of our sources details a fascinating experiment

00:30:20.960 --> 00:30:23.380
using a minimalistic quantum optical reservoir.

00:30:23.900 --> 00:30:26.840
They use just a few atoms trapped inside an optical cavity,

00:30:27.580 --> 00:30:30.240
basically two highly reflective mirrors facing each other.

00:30:30.400 --> 00:30:32.660
-Okay. -And instead of standard discrete measurements,

00:30:33.070 --> 00:30:34.040
they use continuous measurement.

00:30:34.720 --> 00:30:37.320
They let a tiny, continuous stream of light leak out of the cavity,

00:30:37.840 --> 00:30:40.340
and they use a technique called homodyne detection to read it.

00:30:40.560 --> 00:30:43.240
Let's pause on homodyne detection. What exactly is that?

00:30:43.600 --> 00:30:47.180
Simply put, homodyne detection is a way to measure the phase and amplitude

00:30:47.450 --> 00:30:52.240
of an incredibly weak light signal by mixing it with a much stronger known laser beam.

00:30:52.720 --> 00:30:57.020
which they call a local oscillator. By looking at how the weak light interferes with the strong

00:30:57.220 --> 00:31:03.059
light, you can extract the information without destroying the delicate quantum state inside the

00:31:03.100 --> 00:31:07.640
cavity. Okay, so they have this continuous safe stream of information leaking out.

00:31:08.200 --> 00:31:11.340
Where does the feedback actually come in?

00:31:11.639 --> 00:31:12.500
This is the brilliant part.

00:31:12.679 --> 00:31:14.539
They take that classical measurement readout,

00:31:14.950 --> 00:31:16.539
and in real time, they physically

00:31:16.789 --> 00:31:18.899
re-injected as a parameter for the next input.

00:31:19.179 --> 00:31:20.260
We re-injected how?

00:31:20.919 --> 00:31:22.340
Well, for example, if the homodyne detector

00:31:22.639 --> 00:31:25.240
reads a specific phase of light at time t,

00:31:25.700 --> 00:31:31.159
The classical computer instantly uses that information to adjust the intensity of the

00:31:31.220 --> 00:31:35.440
driving laser that is feeding data into the cavity for time t plus 1.

00:31:35.659 --> 00:31:40.720
Oh, so the output of the system essentially becomes part of the environment for the next input.

00:31:41.240 --> 00:31:41.700
Precisely.

00:31:42.139 --> 00:31:46.559
By looping the classical measurement back into the quantum driving field, you create recurrent

00:31:46.779 --> 00:31:49.700
memory holding structure outside the quantum collapse boundary.

00:31:50.019 --> 00:31:50.820
That is wild.

00:31:51.279 --> 00:31:51.480
Right.

00:31:51.960 --> 00:31:54.620
The classical feedback loop acts as an external memory bank.

00:31:55.420 --> 00:31:58.980
It protects the internal quantum system from the destructive back action of measurement.

00:31:59.960 --> 00:32:05.040
The sources show that this feedback enhanced QRC dramatically outperforms systems without

00:32:05.260 --> 00:32:05.420
feedback.

00:32:06.140 --> 00:32:09.680
It remains highly accurate even when environmental noise is introduced.

00:32:10.120 --> 00:32:11.560
So if we look at the whole picture here.

00:32:11.800 --> 00:32:17.700
We've bypassed the input bottleneck and we've solved the amnesia problem using weak measurements and feedback loops.

00:32:18.020 --> 00:32:18.340
We have.

00:32:18.540 --> 00:32:21.900
We finally have a functioning memory retaining quantum reservoir.

00:32:22.680 --> 00:32:27.740
So the next logical question for our clinical listener is, what do we actually do with it?

00:32:28.040 --> 00:32:28.140
Right.

00:32:28.460 --> 00:32:32.140
How does it stack up against the tools we already use to analyze medical data?

00:32:32.620 --> 00:32:37.160
Well, to answer that, we need to quickly survey the current landscape of sequential data processing.

00:32:37.700 --> 00:32:43.620
When medical researchers look at time series data, they usually start with traditional statistical methods like ARMA.

00:32:44.000 --> 00:32:46.139
Right. Auto-regressive integrated moving average.

00:32:46.659 --> 00:32:54.659
Exactly. ARMA is a staple. It looks at the past values in a time series and uses linear regression to predict the next value.

00:32:55.659 --> 00:33:01.799
It is fantastic for data that has clear consistent trends, like, say, a relatively healthy patient's

00:33:01.919 --> 00:33:03.980
resting heart rate over a few hours.

00:33:04.399 --> 00:33:05.899
But ARIMA is inherently linear.

00:33:06.360 --> 00:33:06.460
Yes.

00:33:06.659 --> 00:33:10.840
It completely falls apart when the data becomes highly complex, non-stationary, or chaotic.

00:33:11.200 --> 00:33:15.260
which, let's be honest, describes almost all critical biological data.

00:33:15.539 --> 00:33:15.679
Exactly.

00:33:15.960 --> 00:33:18.200
I mean, the human body is not a linear system.

00:33:18.460 --> 00:33:24.039
A patient going into sudden ventricular fibrillation doesn't follow a neat straight line on a graph.

00:33:24.500 --> 00:33:24.659
Exactly.

00:33:25.519 --> 00:33:28.320
So when a rema fails, the field historically moved to deep learning.

00:33:28.779 --> 00:33:33.179
Specifically, long-short-term memory networks, or LSTMs, and now, transcomers.

00:33:33.500 --> 00:33:37.419
And transcomers are basically the undisputed heavyweights of the AI world right now.

00:33:37.580 --> 00:33:40.899
Oh, they're phenomenal at natural language processing, because they can look at a whole

00:33:41.000 --> 00:33:45.139
paragraph at once and understand the context of every word relative to every other word.

00:33:45.480 --> 00:33:50.019
But transcomers and LSTMs have two massive drawbacks for real-time medical applications.

00:33:50.679 --> 00:33:52.779
They have an insatiable appetite for training data.

00:33:53.100 --> 00:33:56.240
and they require highly intensive back propagation to learn.

00:33:56.799 --> 00:34:01.600
And as we discussed earlier, trying to do back propagation in the quantum realm is an absolute nightmare.

00:34:01.880 --> 00:34:05.400
Doesn't it lead to something the literature calls the barren plateau problem?

00:34:05.620 --> 00:34:13.060
Yes, the vanishing gradient problem, which is basically the bane of standard quantum neural networks or QNNs.

00:34:13.119 --> 00:34:13.700
How does that work?

00:34:13.760 --> 00:34:17.980
When you try to train a quantum circuit using back propagation, you calculate gradients.

00:34:18.500 --> 00:34:22.899
These are the mathematical signals that tell the algorithm which direction to adjust its

00:34:22.950 --> 00:34:23.820
weights to get smarter.

00:34:25.120 --> 00:34:29.220
But as you scale up the number of quibits, those gradients exponentially decay to zero.

00:34:29.710 --> 00:34:34.020
So the signal telling the AI how to improve just vanishes?

00:34:34.200 --> 00:34:34.340
Yes.

00:34:34.850 --> 00:34:38.520
The mathematical training landscape, which you want to look like a nice, sloping valley

00:34:38.679 --> 00:34:41.220
pointing toward the right answer, becomes completely flat.

00:34:41.500 --> 00:34:42.399
a barren plateau.

00:34:43.139 --> 00:34:43.659
Hence the name.

00:34:43.980 --> 00:34:47.399
Right. The network essentially gets stuck in the middle of a desert, completely blind,

00:34:47.770 --> 00:34:50.879
and stops learning entirely no matter how much patient data you feed it.

00:34:51.300 --> 00:34:55.080
But our quantum reservoir doesn't care about barren plateaus at all, does it?

00:34:55.460 --> 00:34:56.000
Not even slightly.

00:34:56.320 --> 00:35:00.120
Because QRC requires zero backpropagation through the quantum circuit.

00:35:00.580 --> 00:35:03.360
Exactly. Remember, the internal physics of the pond are fixed.

00:35:03.700 --> 00:35:08.660
We only train the classical readout sensors at the edge using simple linear regression.

00:35:09.140 --> 00:35:12.560
Because we aren't tweaking the quantum gates, there are no gradients to vanish.

00:35:12.880 --> 00:35:15.420
It completely bypasses the barren plateau problem.

00:35:15.820 --> 00:35:20.160
It does, allowing us to leverage that massive Hilbert space without the training penalty.

00:35:20.619 --> 00:35:25.480
So what kinds of problems is this architecture actually proving to be good at in the real world?

00:35:25.820 --> 00:35:30.740
It excels incredibly well at predicting chaotic non-linear time series.

00:35:32.340 --> 00:35:38.120
Chaos, mathematically speaking, is a system that is deterministic, it follows rules, but

00:35:38.120 --> 00:35:40.240
is exponentially sensitive to initial conditions.

00:35:41.190 --> 00:35:44.380
Think of predicting the weather, or turbulent fluid dynamics in a pipe.

00:35:44.790 --> 00:35:48.880
Or predicting an epileptic seizure based on chaotic EEG brainwave data.

00:35:49.420 --> 00:35:49.660
Exactly.

00:35:50.430 --> 00:35:54.840
And to rigorously measure how well a computing system handles chaos, researchers don't just

00:35:54.930 --> 00:35:55.100
guess.

00:35:55.610 --> 00:35:58.960
They use a highly specific metric called the information processing capacity.

00:35:59.200 --> 00:35:59.940
or IPC.

00:36:00.540 --> 00:36:00.859
IPC.

00:36:01.680 --> 00:36:05.540
OK, so if I'm looking at a paper comparing different AI models for patient outcomes,

00:36:06.359 --> 00:36:09.640
how does an IPC score translate to what I care about?

00:36:10.160 --> 00:36:15.579
Think of IPC as a unified mathematical score of a system's true computational power.

00:36:16.380 --> 00:36:19.140
It evaluates the system by summing up two distinct capabilities.

00:36:19.940 --> 00:36:21.580
First, it's linear memory.

00:36:21.880 --> 00:36:23.220
It's short-term memory capacity.

00:36:23.670 --> 00:36:27.000
Basically, how well can it remember the exact inputs from five steps ago?

00:36:27.200 --> 00:36:28.300
Okay, memory and the second.

00:36:28.680 --> 00:36:30.880
Second is it's non-linear processing capacity.

00:36:31.300 --> 00:36:36.000
How well can it multiply and transform those past inputs to predict complex curves?

00:36:36.920 --> 00:36:41.740
So high IPC means the system can hold on to the past and use it to understand complex twisting futures.

00:36:42.120 --> 00:36:52.360
Yes, and when researchers test QRC on standard chaotic benchmark tests, like the Mackie Glass Time Series or the Santa Fe chaotic data set which are notoriously difficult mathematical curves,

00:36:52.900 --> 00:36:53.960
The results are stunning.

00:36:54.359 --> 00:36:54.460
Really?

00:36:54.920 --> 00:36:55.039
Yeah.

00:36:55.299 --> 00:36:59.500
The literature clearly shows that quantum reservoirs utilizing just five to seven qubits handle

00:36:59.700 --> 00:37:04.000
these chaotic forecasting tasks with an accuracy that equals or surpasses massive classical

00:37:04.400 --> 00:37:04.760
neural networks.

00:37:05.380 --> 00:37:08.480
That is the density of the two to the end Hilbert space at work.

00:37:08.720 --> 00:37:11.400
Just a few atoms doing the work of hundreds of processors.

00:37:11.960 --> 00:37:13.039
It is, however.

00:37:14.280 --> 00:37:18.280
We have to be intellectually honest here and mention a massive development on the classical

00:37:18.700 --> 00:37:22.540
side that our sources highlight because it puts the quantum achievements in perspective.

00:37:22.900 --> 00:37:23.080
Okay.

00:37:23.330 --> 00:37:23.720
What is it?

00:37:23.960 --> 00:37:28.140
It's called next generation reservoir computing or NGRC.

00:37:28.420 --> 00:37:28.840
Next generation.

00:37:29.530 --> 00:37:30.460
What makes it next generation?

00:37:30.890 --> 00:37:32.480
Do they use a different kind of artificial pond?

00:37:33.170 --> 00:37:33.720
This is the twist.

00:37:34.060 --> 00:37:34.300
Yeah.

00:37:34.790 --> 00:37:35.860
They get rid of the pond entirely.

00:37:36.260 --> 00:37:36.540
Wait, what?

00:37:36.880 --> 00:37:37.000
Yeah.

00:37:37.700 --> 00:37:42.500
NGRC actually strips away the random physically fixed matrices.

00:37:43.100 --> 00:37:45.260
It is a completely recurrence-free approach.

00:37:46.220 --> 00:37:48.940
Instead of feeding data into a dynamic network,

00:37:49.320 --> 00:37:53.000
it relies on purely mathematical nonlinear vector autoregression.

00:37:53.320 --> 00:37:54.200
Okay, that's a mouthful.

00:37:54.200 --> 00:37:57.400
It explicitly calculates polynomial features of delayed inputs.

00:37:57.800 --> 00:38:00.000
So instead of letting the water do the math naturally,

00:38:00.160 --> 00:38:04.880
they found a way to just do the highly complex math directly and efficiently on a classical computer.

00:38:05.340 --> 00:38:05.520
Yes.

00:38:06.400 --> 00:38:07.440
And the results are staggering.

00:38:08.480 --> 00:38:14.420
NGRC has been shown to train up to 10,000 times faster than competing deep learning approaches

00:38:14.650 --> 00:38:17.560
on complex spatial temporal chaos tasks.

00:38:18.100 --> 00:38:19.900
10,000 times faster.

00:38:20.030 --> 00:38:22.300
And it requires vastly less training data.

00:38:22.720 --> 00:38:24.480
to achieve the exact same accuracy.

00:38:24.880 --> 00:38:28.680
Wait a minute, if classical next-gen RC is 10,000 times faster,

00:38:29.340 --> 00:38:33.100
requires less data and doesn't need super-cooled quantum chips,

00:38:33.780 --> 00:38:36.620
why on earth do we need quantum reservoir computing at all?

00:38:36.820 --> 00:38:39.200
That is the multi-million dollar question in the field right now.

00:38:39.280 --> 00:38:39.520
I bet.

00:38:39.860 --> 00:38:43.560
NGRC is profoundly important because it separates the conceptual power of

00:38:43.660 --> 00:38:45.900
reservoir computing from its physical implementation.

00:38:46.460 --> 00:38:50.600
It proves mathematically that the core idea expanding inputs into a massive feature space

00:38:51.060 --> 00:38:53.880
and only training a linear readout works brilliantly for chaos.

00:38:54.220 --> 00:38:55.400
So it validates the whole theory.

00:38:55.800 --> 00:38:58.940
Right. And the speculation now and the frontier where researchers are heading

00:38:59.380 --> 00:39:01.660
is whether we can build a quantum equivalent of NGRC.

00:39:01.880 --> 00:39:02.140
Oh.

00:39:02.780 --> 00:39:06.680
Taking the mathematical purity of next-gen RC and supercharging it with quantum dimensions.

00:39:07.300 --> 00:39:12.780
Exactly. Can we use quantum entanglement to generate those high-order, non-linear polynomial

00:39:12.980 --> 00:39:18.420
features even more efficiently than classical vector autoregression can? If we can map the

00:39:18.540 --> 00:39:21.060
NGRC algorithms onto a quantum substrate,

00:39:21.500 --> 00:39:24.620
the efficiency gains could be literally unimaginable.

00:39:25.040 --> 00:39:28.000
Okay, this is fascinating theory, but let's bring it back to the clinic.

00:39:28.160 --> 00:39:28.440
That's good.

00:39:28.900 --> 00:39:33.660
Because predicting a mathematical chaotic attractor on a laboratory computer is one thing.

00:39:34.180 --> 00:39:40.020
Predicting the chaotic biological rhythms of a human patient in real time in a hospital bed is

00:39:40.020 --> 00:39:41.060
an entirely different beast.

00:39:41.500 --> 00:39:46.720
Let's move into the heart of entangled health, precision medicine and biological noise.

00:39:47.120 --> 00:39:48.480
It is the ultimate testing ground.

00:39:48.880 --> 00:39:52.800
When you look at real clinical data, say an electroencephalogram measuring brainways

00:39:52.900 --> 00:39:55.800
or an electrocardiogram measuring the heart, it is incredibly messy.

00:39:56.320 --> 00:39:57.020
It's not clean.

00:39:57.600 --> 00:39:59.580
isolated laboratory data.

00:39:59.820 --> 00:40:03.920
Biological data is plagued by what you called earlier non-stationary noise.

00:40:04.240 --> 00:40:05.580
I mean, the human body isn't static.

00:40:05.690 --> 00:40:08.540
A patient's baseline heart rate shifts when they shift in bed,

00:40:08.570 --> 00:40:11.680
when they take a deep breath, when their circadian rhythm changes at 2 a.m.

00:40:12.080 --> 00:40:13.580
or if they just metabolize a drug differently.

00:40:14.240 --> 00:40:18.260
Exactly. Respiration, movement, digestion, it all bleeds into the data.

00:40:18.540 --> 00:40:18.620
Yeah.

00:40:19.020 --> 00:40:23.300
Furthermore, there are complex, highly non-linear cross-channel dependencies.

00:40:23.710 --> 00:40:24.680
Give me an example of that.

00:40:25.040 --> 00:40:30.980
Well, a subtle spike in one biological marker, say a specific protein level, might only be

00:40:31.100 --> 00:40:35.200
medically relevant if it occurs exactly three seconds after a slight dip in blood oxygen.

00:40:35.800 --> 00:40:38.240
But only while the patient is in REM sleep.

00:40:38.620 --> 00:40:38.840
Wow.

00:40:39.460 --> 00:40:44.920
And classical linear memory models like ARIMA just look at that and see pure unreadable static.

00:40:45.280 --> 00:40:45.600
They do.

00:40:46.000 --> 00:40:49.140
They struggle to capture those deep hidden temporal correlations.

00:40:49.940 --> 00:40:54.420
And this is exactly where the massive feature expressivity of the quantum Hilbert space shines.

00:40:55.040 --> 00:40:56.920
Because it can see the whole picture at once.

00:40:57.140 --> 00:40:57.260
Right.

00:40:57.840 --> 00:41:03.120
QRC, by mapping that noisy data into thousands of dimensions simultaneously, can capture correlations

00:41:03.250 --> 00:41:07.600
and patterns in biological rhythms that simply slip through the net of classical memory.

00:41:07.940 --> 00:41:10.960
Let's give the listener some concrete examples directly from the sources we reviewed.

00:41:11.310 --> 00:41:14.820
We have studies showing reservoir computing being used to model the highly non-linear

00:41:15.040 --> 00:41:17.440
relationships in the penicillin fermentation process.

00:41:18.130 --> 00:41:20.420
That is a great industrial biology example.

00:41:21.020 --> 00:41:24.619
Fermentation is an incredibly delicate, complex temporal process.

00:41:25.280 --> 00:41:30.839
Small deviations in temperature or nutrient mix early on yield completely ruined batches

00:41:31.140 --> 00:41:31.740
days later.

00:41:32.320 --> 00:41:34.400
You have to predict the chaos before it manifests.

00:41:34.700 --> 00:41:34.800
Right.

00:41:35.040 --> 00:41:39.579
In another source, extensively details using quantum reservoirs for predicting molecular

00:41:39.880 --> 00:41:41.380
properties to accelerate drug discovery.

00:41:42.060 --> 00:41:43.940
We even have the Core-D's project.

00:41:44.160 --> 00:41:46.360
Ah, yes. Core-D's.

00:41:46.640 --> 00:41:52.080
That is a major collaboration between Ludwig Maximilian University in Munich and the pharmaceutical giant Merck.

00:41:52.280 --> 00:41:53.500
Dell-M-U and Merck, right?

00:41:53.600 --> 00:41:58.440
Yeah. They are explicitly exploring how quantum machine learning, specifically these reservoir techniques,

00:41:59.080 --> 00:42:02.260
can model the complex binding affinities of new drug compounds.

00:42:02.780 --> 00:42:06.260
But the case study that really stood out to me was the Clinical Trial Prediction Project.

00:42:06.500 --> 00:42:11.540
the sources outline a massive collaboration involving Quera, which is a leading quantum hardware company,

00:42:12.160 --> 00:42:17.540
along with Merck, Amgen and Deloitte. They are using these models to forecast clinical trial outcomes.

00:42:22.300 --> 00:42:31.380
This is one of these moments where these prompts mix something up. Let's take a break and re-render

00:42:31.380 --> 00:42:32.040
the whole session.

00:42:33.300 --> 00:42:36.360
Right, and we are actually seeing the groundwork for this being laid right now.

00:42:36.540 --> 00:42:37.860
Take the quartz project, for example.

00:42:38.040 --> 00:42:39.640
Oh yeah, QU-A-R-D-S, right.

00:42:39.880 --> 00:42:41.660
You announced exactly like the mineral, quartz.

00:42:42.060 --> 00:42:43.180
Exactly, yeah, like the rock.

00:42:44.260 --> 00:42:47.500
So this is a huge three-year funded collaboration.

00:42:48.120 --> 00:42:52.760
It's anchored by LMU Munich and the pharma company Merck-KGA-Darmstadt.

00:42:53.260 --> 00:42:57.060
And their goal there, which I think is just fascinating, isn't just to write theoretical

00:42:57.340 --> 00:42:57.920
physics papers.

00:42:58.400 --> 00:43:01.240
they are actively trying to figure out, you know,

00:43:01.520 --> 00:43:03.980
which specific data structures actually fit

00:43:04.579 --> 00:43:05.220
reservoir computing.

00:43:05.319 --> 00:43:07.880
>> Right, they're hunting for data that has these really deep,

00:43:08.200 --> 00:43:09.140
complex, hidden correlations.

00:43:09.619 --> 00:43:11.279
>> Yeah, like systems, biology effects,

00:43:11.640 --> 00:43:15.000
or forecasting clinical trial outcomes,

00:43:15.500 --> 00:43:18.119
or, and this is wild, using human voice

00:43:18.660 --> 00:43:22.400
as a vocal biomarker to detect a patient's overall quality of life.

00:43:22.440 --> 00:43:23.279
>> Wait, voice analysis?

00:43:23.760 --> 00:43:25.279
Like just listening to someone talk.

00:43:25.420 --> 00:43:27.180
>> Yeah, actually within that frame,

00:43:27.740 --> 00:43:31.640
Merck and Ed Zürich just finalized a master's thesis as a proof of concept,

00:43:32.090 --> 00:43:34.200
and the results were really promising. They worked with this

00:43:34.540 --> 00:43:36.220
German startup called Peak Profiling.

00:43:36.560 --> 00:43:39.060
Peak profiling, okay. What makes them different from

00:43:39.500 --> 00:43:43.780
standard voice analysis? Well, they go way beyond just, you know, physical acoustic markers.

00:43:44.700 --> 00:43:47.900
They're not just measuring how loud a cough is or the pitch of your voice.

00:43:48.020 --> 00:43:53.000
They develop medical voice biomarkers using AI-driven voice analysis, but they use musicology.

00:43:53.599 --> 00:43:57.280
Musicology, you mean like mapping out the harmonic structure of how someone speaks?

00:43:57.720 --> 00:43:58.000
Exactly.

00:43:59.200 --> 00:44:01.480
They use the rhythmic patterns and harmonies of speech.

00:44:02.000 --> 00:44:06.960
to detect medical conditions and they do it objectively, really quickly, and at massive scale.

00:44:07.480 --> 00:44:13.480
That is just... wow. I mean, feeding the structural complexity of musicology into a quantum reservoir

00:44:13.700 --> 00:44:19.060
just feels like a perfect conceptual fit. It really does. But, you know, the quartz project is also

00:44:19.300 --> 00:44:20.440
tackling hard medical imaging.

00:44:21.480 --> 00:44:25.280
There's another collaboration in the project with Unibon where they're doing this explorative

00:44:25.660 --> 00:44:25.800
study.

00:44:26.040 --> 00:44:28.680
Right, where they're developing that PBML framework.

00:44:28.920 --> 00:44:30.320
Yeah, physics-based machine learning.

00:44:31.040 --> 00:44:36.620
So Unibon is developing this framework to extract quantitative, physically meaningful parameters

00:44:37.000 --> 00:44:39.680
from CES to MRI data.

00:44:40.080 --> 00:44:42.740
CES being chemical exchange saturation transfer, right?

00:44:43.080 --> 00:44:43.320
Exactly.

00:44:43.820 --> 00:44:48.300
Instead of just relying on the old qualitative contrast metrics, they are baking physics into

00:44:48.300 --> 00:44:50.840
the machine learning to get real hard numbers.

00:44:51.340 --> 00:44:54.640
And just as a side note to show how rigorously they are testing this,

00:44:55.020 --> 00:44:59.960
They actually take these trained PBML models and apply them to Phantom's CES measurements.

00:45:00.339 --> 00:45:05.240
Right, Phantoms. Like the standard calibration objects they put in MRI machines rather than

00:45:05.559 --> 00:45:11.680
human patients. Yeah, and they use Phantoms with independently measured known pH levels and T1 and T2

00:45:11.819 --> 00:45:12.660
relaxation times.

00:45:13.440 --> 00:45:18.299
So they do this to test whether this physics-based approach yields more robust and interpretable

00:45:18.440 --> 00:45:23.640
estimates than conventional CES metrics like magnetization transfer ratio asymmetry.

00:45:23.839 --> 00:45:27.960
Which is quite a mouthful, but yeah, it is a perfect example of moving away from qualitative

00:45:28.260 --> 00:45:31.220
guesswork and into high-dimensional quantitative analysis.

00:45:31.720 --> 00:45:35.380
And actually speaking of high-dimensional analysis, we have to talk about that recent

00:45:35.440 --> 00:45:36.400
proof of concept paper.

00:45:36.600 --> 00:45:40.940
The one with Quara, Amgen, Deloitte, and again, Mercagie, A. Darmstadt.

00:45:41.300 --> 00:45:45.580
Oh yeah, and we really need to correct a very common misconception about this specific paper.

00:45:46.080 --> 00:45:49.200
Because people see those massive pharmaceutical names together,

00:45:49.560 --> 00:45:52.040
and they immediately assume it was a clinical trial application.

00:45:52.780 --> 00:45:53.500
which was not.

00:45:53.780 --> 00:45:54.620
No, absolutely not.

00:45:54.790 --> 00:46:00.220
It was a rigorous proof of concept designed to predict molecular activity using molecular

00:46:00.720 --> 00:46:01.020
fingerprints.

00:46:01.600 --> 00:46:05.700
And there's a really fun bit of irony buried in the methodology here, too.

00:46:05.940 --> 00:46:06.200
Oh, yeah.

00:46:06.340 --> 00:46:06.860
I love this part.

00:46:07.080 --> 00:46:11.780
So to train the system, they used this black box data set from the Merck Molecular Activity

00:46:12.080 --> 00:46:15.420
Challenge, which actually originates from the other Merck, the pharmaceutical company

00:46:15.640 --> 00:46:16.220
based in the US.

00:46:16.580 --> 00:46:16.720
Right.

00:46:16.840 --> 00:46:20.440
Not their collaborators in Darmstadt, just a great little twist of industry fate.

00:46:20.740 --> 00:46:21.000
Mm-hmm.

00:46:21.440 --> 00:46:26.079
But, if we connect these two points, predicting molecular activity is really just the necessary

00:46:26.480 --> 00:46:27.140
baseline proof.

00:46:27.710 --> 00:46:27.839
Exactly.

00:46:28.500 --> 00:46:32.740
If a quantum reservoir can successfully handle the incredibly complex, multi-dimensional

00:46:33.119 --> 00:46:38.180
fingerprints of chemical molecules, then the logical future step, which the entire quartz

00:46:38.339 --> 00:46:40.380
project is looking toward, is scaling that up.

00:46:40.600 --> 00:46:46.560
applying that exact same mechanism to massive correlated patient trajectories in actual medicine.

00:46:46.880 --> 00:46:50.740
Right. Comparing the quantum approach against classical machine learning approaches,

00:46:51.320 --> 00:46:56.600
finding the synergies between them, and really identifying the specific types of clinical data

00:46:56.690 --> 00:46:59.880
that most benefit from being thrown into that quantum pond.

00:47:00.240 --> 00:47:00.720
(gentle music)

00:47:03.280 --> 00:47:07.080
clinical trial prediction is arguably one of the most valuable use cases in

00:47:07.080 --> 00:47:11.060
modern medicine. When you run a trial you are tracking incredibly complex

00:47:11.520 --> 00:47:16.380
patient trajectories over months or years. You are trying to forecast efficacy.

00:47:17.480 --> 00:47:19.060
You know, is the drug actually working?

00:47:19.820 --> 00:47:23.340
and trying to predict severe adverse events before they happen.

00:47:23.620 --> 00:47:31.060
Great. And if a quantum model can see the hidden correlation that a patient is heading towards liver toxicity weeks before a classical model can,

00:47:31.610 --> 00:47:34.860
well that saves lives and it saves billions of dollars and failed trials.

00:47:35.160 --> 00:47:37.360
Absolutely. But here is where I really need to push back.

00:47:37.600 --> 00:47:40.280
Let's talk about the reality of medical data volume.

00:47:40.700 --> 00:47:40.860
Okay.

00:47:41.980 --> 00:47:43.720
Here is where I really need to push back.

00:47:43.960 --> 00:47:46.060
Let's talk about the reality of medical data volume.

00:47:46.840 --> 00:47:49.920
Classical models like transformers need billions of data points.

00:47:50.410 --> 00:47:53.720
Earlier, we talked about how transformers need billions of data points.

00:47:54.100 --> 00:47:59.100
We hear about large language models essentially reading the entire internet to learn how to speak.

00:47:59.410 --> 00:48:00.540
Yes, huge data sets.

00:48:01.040 --> 00:48:07.080
But in clinical trials or in personalized precision medicine, we simply do not have billions of data points.

00:48:07.780 --> 00:48:10.480
A phase two clinical trial might only have 50 patients.

00:48:11.200 --> 00:48:15.200
Or if I am trying to build a personalized heart monitor for one specific patient,

00:48:15.680 --> 00:48:18.580
my entire data set is just that one individual.

00:48:20.020 --> 00:48:23.460
How does a massive computational system handle small data?

00:48:23.980 --> 00:48:28.380
This is perhaps the most crucial advantage of reservoir computing, both classical and

00:48:28.500 --> 00:48:30.720
quantum, for the entire healthcare sector.

00:48:31.500 --> 00:48:36.280
QRC demonstrates incredible unparalleled robustness on small data sets.

00:48:36.880 --> 00:48:39.720
How is that possible? AI always needs big data.

00:48:40.100 --> 00:48:43.000
Deep learning needs big data because it has to train the internal layers.

00:48:51.480 --> 00:48:53.480
If you have a network with a billion empty weights,

00:48:53.870 --> 00:48:58.540
you need a billion examples to teach the network how to extract basic features like what is a heartbeat.

00:48:58.900 --> 00:48:59.680
Right, that makes sense.

00:48:59.810 --> 00:49:02.180
But remember, in a reservoir, you don't train the inside.

00:49:02.600 --> 00:49:05.940
The natural physics of the quantum pond extracts the features for you instantly.

00:49:06.340 --> 00:49:09.700
Oh, so you just need enough data to fit the linear readout at the edge.

00:49:10.120 --> 00:49:14.920
Precisely, you just need enough data to run a simple linear regression.

00:49:15.920 --> 00:49:19.840
It requires a minuscule amount of training data to stabilize mathematically.

00:49:19.910 --> 00:49:20.660
It is amazing.

00:49:20.850 --> 00:49:23.020
It makes it perfect for personalized precision medicine,

00:49:23.290 --> 00:49:26.200
where the dataset is fundamentally limited to, well, you.

00:49:26.700 --> 00:49:30.020
And importantly, it makes it ideal for edge computing devices.

00:49:30.800 --> 00:49:33.700
Just to clarify, edge computing means the processing is happening locally,

00:49:34.160 --> 00:49:37.900
right there on the device, rather than beaming the data up to a cloud server.

00:49:38.180 --> 00:49:42.280
Yes. Think of a wearable medical monitor, like an advanced smartwatch,

00:49:42.740 --> 00:49:43.960
or an implanted pacemaker.

00:49:44.400 --> 00:49:51.640
Imagine an onboard device that learns the specific, chaotic, highly individualized baseline of your heart in real time.

00:49:51.920 --> 00:49:53.020
Okay, I can picture that.

00:49:53.279 --> 00:49:59.440
Because the training algorithm is just a simple mathematical inversion of the readout layer, it uses almost zero battery power.

00:50:00.120 --> 00:50:05.420
It doesn't need to transmit your private health data via Wi-Fi to a massive server farm to run back propagation.

00:50:06.270 --> 00:50:09.560
It processes the chaos locally, instantly, right on the edge.

00:50:14.640 --> 00:50:18.280
Okay, I love this vision and the physics is not wrong.

00:50:18.710 --> 00:50:20.800
The rest of computing is structurally...

00:50:21.099 --> 00:50:25.440
Well suited to low power edge inference, but let's be precise.

00:50:26.160 --> 00:50:30.260
What exists today is a proof of concept hardware in controlled lab settings.

00:50:30.559 --> 00:50:35.299
So the variable revolution is a direction and not a product yet.

00:50:35.760 --> 00:50:36.279
But let's go.

00:50:37.880 --> 00:50:41.799
That sounds like an absolute revolution for remote health care and patient privacy.

00:50:42.339 --> 00:50:42.799
But...

00:50:43.680 --> 00:50:44.580
Let's do a reality check.

00:50:44.900 --> 00:50:47.820
We have been talking about QRC for the last, what, 30 minutes?

00:50:48.060 --> 00:50:51.500
Like, it is a finished product already saving lives in hospitals.

00:50:51.880 --> 00:50:52.320
Fair point.

00:50:52.560 --> 00:50:54.760
Is quantum advantage actually proven here?

00:50:55.140 --> 00:50:57.040
Or are we swimming in theoretical height?

00:50:57.580 --> 00:51:01.420
It is vital to provide an honest, rigorous assessment here.

00:51:02.240 --> 00:51:04.480
And the literature is very clear about the caveats.

00:51:05.280 --> 00:51:08.540
The exponential dimension of the Hilbert space is mathematically real.

00:51:09.060 --> 00:51:13.360
The theoretical processing power is real, but real-world quantum hardware today

00:51:13.960 --> 00:51:18.080
exists in what we call the NISQ-era noisy intermediate-scale quantum devices.

00:51:19.060 --> 00:51:20.880
And they suffer from significant physical limitations.

00:51:21.280 --> 00:51:21.420
Right.

00:51:21.570 --> 00:51:22.660
They are incredibly fragile.

00:51:22.850 --> 00:51:23.240
They are.

00:51:23.430 --> 00:51:27.360
We have shot noise, which arises from the statistical probabilistic nature of measuring

00:51:27.600 --> 00:51:28.100
quantum states.

00:51:28.790 --> 00:51:32.740
We have gate errors, where the microwave pulses trying to manipulate the atoms are

00:51:32.860 --> 00:51:33.700
slightly imprecise.

00:51:33.860 --> 00:51:35.260
And coherence times, right?

00:51:35.660 --> 00:51:36.780
Most critically, yes.

00:51:37.440 --> 00:51:38.640
Finite coherence times.

00:51:39.820 --> 00:51:44.040
The quibits lose their quantum entanglement to the surrounding environment very, very quickly.

00:51:44.960 --> 00:51:51.260
Furthermore, if you're completely honest, a massive, optimized, classical reservoir running on a supercomputer

00:51:51.820 --> 00:51:57.100
can eventually approximate the performance of a small, noisy quantum reservoir today,

00:51:57.680 --> 00:52:01.660
even if the classical machine requires more physical space and a higher energy cost.

00:52:02.280 --> 00:52:04.580
So true, undeniable quantum advantage.

00:52:05.100 --> 00:52:10.420
where a quantum reservoir does something that a classical computer physically, mathematically, cannot do,

00:52:11.020 --> 00:52:14.800
has not been definitively proven for real-world biological datasets yet.

00:52:15.000 --> 00:52:19.780
Correct. The noise of the physical hardware often threatens to drown out the computational signal.

00:52:19.960 --> 00:52:21.540
And this was one of my favorite parts of the research.

00:52:21.760 --> 00:52:25.440
Some researchers are finding that this hardware noise might not be entirely an enemy right.

00:52:25.480 --> 00:52:27.200
Oh, it is a brilliant twist of perspective.

00:52:27.620 --> 00:52:29.380
How could noise possibly be a good thing?

00:52:29.840 --> 00:52:33.000
Well, think back to the pond. Remember when we talked about fading memory?

00:52:33.600 --> 00:52:35.140
The pond has to calm down.

00:52:35.690 --> 00:52:36.940
The ripples have to die out,

00:52:36.970 --> 00:52:38.460
where the system becomes too chaotic.

00:52:38.880 --> 00:52:40.660
- Yes, the fading memory property is essential.

00:52:41.260 --> 00:52:43.840
- But in a perfectly isolated, mathematically ideal

00:52:44.000 --> 00:52:46.680
quantum system, the evolution is unitary.

00:52:47.230 --> 00:52:48.600
The waves never lose energy.

00:52:49.010 --> 00:52:51.100
The information bounces around infinitely.

00:52:51.410 --> 00:52:52.040
It never fades.

00:52:52.600 --> 00:52:53.380
- We, really?

00:52:53.600 --> 00:52:55.400
- Yeah, a perfect quantum computer

00:52:55.540 --> 00:52:57.520
actually makes for a terrible reservoir,

00:52:58.080 --> 00:52:59.320
because it remembers everything forever.

00:52:59.620 --> 00:53:02.200
- Exactly, so to get a quantum system

00:53:02.200 --> 00:53:03.100
to act like a reservoir.

00:53:03.600 --> 00:53:07.040
actually require dissipation. You need some loss. You need the environment to

00:53:07.100 --> 00:53:11.640
bleed some energy away. Exactly. Researchers are discovering that tunable noise

00:53:11.960 --> 00:53:15.740
controlling the natural quantum decoherence of the hardware can actually be

00:53:15.980 --> 00:53:20.240
harnessed as a computational resource. The hardware's inability to stay

00:53:20.400 --> 00:53:24.140
perfectly quantum provides the exact physical mechanism needed for fading

00:53:24.300 --> 00:53:28.520
memory. They are literally turning a hardware bug into a machine learning

00:53:28.720 --> 00:53:28.880
future.

00:53:29.720 --> 00:53:34.099
Using the universe's natural decay to help the machine forget just enough to be useful.

00:53:34.329 --> 00:53:35.579
That is incredible engineering.

00:53:36.000 --> 00:53:36.480
It really is.

00:53:37.400 --> 00:53:42.299
So, if this technology is overcoming its limitations and the horizon is approaching,

00:53:43.039 --> 00:53:44.920
we need to map out the field for our listeners.

00:53:45.840 --> 00:53:47.780
Who is actually building these machines?

00:53:48.660 --> 00:53:52.380
What does a physical quantum reservoir actually look like in a laboratory?

00:53:52.740 --> 00:53:55.500
Well, the field map in the sources is incredibly diverse,

00:53:55.760 --> 00:54:00.160
because as we established, any complex dynamical system can act as a reservoir.

00:54:00.800 --> 00:54:02.640
Researchers are trying literally everything.

00:54:03.100 --> 00:54:04.440
Let's start with the industry giants.

00:54:04.930 --> 00:54:08.260
The sources heavily contrast the different approaches IBM is taking.

00:54:08.440 --> 00:54:13.360
Yes, IBM is a fascinating case study because they have deep industry leading expertise in

00:54:13.540 --> 00:54:16.360
both classical neuromorphic chips and pure quantum systems.

00:54:16.690 --> 00:54:18.280
What are they doing on the neuromorphic side?

00:54:18.440 --> 00:54:22.640
On the neuromorphic side, which is basically advanced classical physical reservoir computing,

00:54:23.330 --> 00:54:27.740
they are building chips that mimic brain architecture and are staggeringly energy efficient.

00:54:28.120 --> 00:54:33.320
We're talking about utilizing just 70 milliwatts to run a network of 1 million artificial neurons.

00:54:33.600 --> 00:54:38.640
70 milliwatts. A standard incandescent light bulb is what 60 watts, so IBM is running a million

00:54:38.770 --> 00:54:43.440
neuron brain on a fraction of the power of a desk lamp. It is incredibly efficient,

00:54:43.840 --> 00:54:46.560
perfect for those edge computing wearable devices we discussed.

00:54:47.460 --> 00:54:50.620
But on the quantum side, IBM is taking a completely different route.

00:54:51.220 --> 00:54:55.320
They are building massive, gate-based, superconducting quantum computers.

00:54:55.580 --> 00:54:58.000
The ones that look like giant gold chandeliers.

00:54:58.180 --> 00:54:58.460
Exactly.

00:54:58.680 --> 00:55:03.820
They have high quibbit counts and leading coherence times, but those systems require extreme,

00:55:04.480 --> 00:55:04.960
cooling infrastructure.

00:55:05.600 --> 00:55:10.080
They are built inside massive cryostats that drop the temperature down to near absolute

00:55:10.260 --> 00:55:10.480
zero.

00:55:10.740 --> 00:55:15.600
You are definitely not putting an absolute zero cryostat in a wearable heart monitor anytime

00:55:16.010 --> 00:55:16.120
soon.

00:55:16.300 --> 00:55:16.900
No, you are not.

00:55:17.310 --> 00:55:18.960
But IBM isn't the only player.

00:55:19.630 --> 00:55:24.020
The sources explicitly mention Quara, the company involved in the clinical trial prediction.

00:55:24.740 --> 00:55:26.079
They use an entirely different platform.

00:55:27.339 --> 00:55:29.839
Analog, Rideberg-Adam quantum computers.

00:55:30.119 --> 00:55:32.160
- Okay, how does a Rideberg-Adam computer work?

00:55:32.299 --> 00:55:33.079
That sounds sci-fi.

00:55:33.380 --> 00:55:34.760
- Instead of superconducting circuits,

00:55:34.990 --> 00:55:36.339
they use highly focused lasers

00:55:36.480 --> 00:55:38.119
to trap individual neutral atoms,

00:55:38.520 --> 00:55:40.339
usually rubidium in a complete vacuum.

00:55:40.900 --> 00:55:43.520
By shining precisely tuned lasers at these atoms,

00:55:44.059 --> 00:55:45.520
they excite their outermost electron

00:55:46.220 --> 00:55:47.599
far, far away from the nucleus,

00:55:48.020 --> 00:55:49.440
turning them into ride-burg atoms.

00:55:49.720 --> 00:55:50.359
And what does that do?

00:55:50.760 --> 00:55:52.000
When atoms are in this state,

00:55:52.319 --> 00:55:55.020
they interact with each other over relatively huge distances

00:55:55.599 --> 00:55:57.200
via strong dipole forces.

00:55:58.440 --> 00:56:01.000
That interaction creates the entanglement,

00:56:01.319 --> 00:56:02.940
the whispering network of our pond,

00:56:03.579 --> 00:56:05.400
without needing any physical wires or circuits.

00:56:05.900 --> 00:56:07.619
So it's a pond made of laser-trapped atoms.

00:56:07.839 --> 00:56:08.200
That's amazing.

00:56:09.500 --> 00:56:13.480
What else? The source is mentioned nuclear magnetic resonance or NMR?

00:56:13.740 --> 00:56:18.180
Yes, NMR spin ensembles. This is very robust approach that actually operates at room temperature.

00:56:18.520 --> 00:56:21.620
It uses the natural nuclear spin dynamics of molecules in a liquid.

00:56:21.960 --> 00:56:28.020
Wait, nuclear magnetic resonance isn't that the exact same physics behind an MRI machine in a hospital?

00:56:28.260 --> 00:56:34.320
It is exactly the same underlying physics. In an MRI, a massive magnet aligns the spin of the

00:56:34.380 --> 00:56:36.800
hydrogen protons in the water inside your body.

00:56:37.100 --> 00:56:42.180
Then, radiofrequency pulses knock those spins out of alignment, and sensors measure the radio

00:56:42.250 --> 00:56:44.700
waves the protons emit as they relax back to baseline.

00:56:45.010 --> 00:56:48.140
And that relaxation data creates the image of your brain or knee.

00:56:48.420 --> 00:56:48.580
Exactly.

00:56:49.120 --> 00:56:50.240
So how does that become a computer?

00:56:50.740 --> 00:56:56.000
In an NMR reservoir, a liquid containing specific molecules is placed in a magnetic field.

00:56:56.930 --> 00:57:01.020
The input data is encoded into radiofrequency pulses that hit the liquid.

00:57:01.500 --> 00:57:06.440
The complex interacting physics of millions of molecules tumbling and transferring spin

00:57:06.620 --> 00:57:09.620
states to one another acts as the nonlinear reservoir.

00:57:10.200 --> 00:57:15.060
And then you then use a sensor coil to read the total magnetization of the liquid as your

00:57:15.200 --> 00:57:15.400
output.

00:57:15.800 --> 00:57:21.440
You are literally dropping a mathematical stone into a physical vial of liquid and reading

00:57:21.580 --> 00:57:22.560
the magnetic ripples.

00:57:23.200 --> 00:57:24.520
That is beautifully literal.

00:57:24.820 --> 00:57:25.060
It is.

00:57:25.500 --> 00:57:27.500
and there are even more elegant approaches.

00:57:34.860 --> 00:57:38.860
Some teams are using superconducting circuits with parametrically coupled oscillators.

00:57:39.760 --> 00:57:43.100
Instead of using distinct binary quibits, which are just zeros and ones,

00:57:43.800 --> 00:57:45.580
They use quantum harmonic oscillators.

00:57:45.860 --> 00:57:48.100
Think of them like quantum pendulums swinging back and forth.

00:57:48.840 --> 00:57:48.960
Exactly.

00:57:49.320 --> 00:57:53.700
The amplitude and phase of the pendulum can represent many different states at once, utilizing

00:57:53.820 --> 00:57:55.220
what physicists call "fox" states.

00:57:55.920 --> 00:57:57.360
The information density is incredible.

00:57:57.700 --> 00:57:58.660
How dense are we talking?

00:57:58.920 --> 00:58:02.380
Well, when paper highlighted in the source is demonstrated that by using just two coupled

00:58:02.580 --> 00:58:03.400
quantum oscillators,

00:58:03.700 --> 00:58:07.500
they could represent a network of 81 distinct artificial neurons.

00:58:07.960 --> 00:58:10.480
Two physical pendulums doing the work of 81 nodes.

00:58:10.740 --> 00:58:10.839
Yes.

00:58:11.430 --> 00:58:15.700
And they achieved 99% accuracy on challenging benchmark classification tasks.

00:58:16.440 --> 00:58:20.859
To get that same accuracy classically, you would need at least 24 classical oscillators.

00:58:21.460 --> 00:58:26.060
The density of information encoding in continuous variable quantum systems is remarkable.

00:58:26.460 --> 00:58:26.680
Wow.

00:58:27.099 --> 00:58:29.060
And lastly, there are optical approaches right.

00:58:29.740 --> 00:58:30.839
- Reservoir is built out of light.

00:58:31.059 --> 00:58:32.720
- Yes, all optical echo state networks.

00:58:33.200 --> 00:58:34.440
These are incredibly promising

00:58:34.539 --> 00:58:36.240
for high-speed telecon and edge devices.

00:58:36.770 --> 00:58:38.799
They utilize phenomena like stimulated

00:58:39.160 --> 00:58:40.720
brilliant scattering and optical fibers.

00:58:40.960 --> 00:58:41.559
- Brilliant scattering.

00:58:42.240 --> 00:58:44.960
- Essentially, the input light wave creates an acoustic wave,

00:58:44.990 --> 00:58:47.079
a sound wave inside the glass fiber,

00:58:47.599 --> 00:58:49.180
which then scatters subsequent light waves

00:58:49.180 --> 00:58:51.079
in highly complex nonlinear ways.

00:58:51.650 --> 00:58:53.760
It allows for measurement-free nonlinear activation,

00:58:54.299 --> 00:58:56.319
drastically reducing energy consumption

00:58:56.480 --> 00:58:58.440
compared to traditional software-based methods.

00:58:58.900 --> 00:59:03.740
Light is incredibly fast, it generates very little heat, and it naturally multiplexes information.

00:59:04.400 --> 00:59:10.299
So we have rubidium atoms trapped by lasers, liquid spins acting like MRI brains,

00:59:11.000 --> 00:59:14.359
superconducting pendulums, and sound waves inside optical fibers,

00:59:14.859 --> 00:59:16.740
all racing to be the best physical pond.

00:59:17.119 --> 00:59:17.500
Quite the lineup.

00:59:17.960 --> 00:59:20.299
But as an engineer looking at clinical applications,

00:59:20.860 --> 00:59:26.740
How do we actually bridge the gap from these massive, delicate lab setups to real-world

00:59:26.920 --> 00:59:28.440
medical use in the next five years?

00:59:29.120 --> 00:59:33.760
The consensus trajectory to the clinic, at least in the near term, relies on a hybrid

00:59:33.880 --> 00:59:34.160
architecture.

00:59:35.320 --> 00:59:38.540
Hybrid quantum classical reservoir computing, or HQCRC.

00:59:39.300 --> 00:59:40.380
bringing the two worlds together.

00:59:40.540 --> 00:59:41.020
- Precisely.

00:59:41.080 --> 00:59:43.580
You don't ask the quantum computer to do everything.

00:59:43.880 --> 00:59:46.540
You let the quantum hardware do the one specific thing it does

00:59:46.780 --> 00:59:48.000
better than anything else in the universe.

00:59:48.740 --> 00:59:50.100
Instantly generate a massive,

00:59:50.460 --> 00:59:52.800
wildly complex, non-linear feature space.

00:59:53.400 --> 00:59:55.380
- But you do not rely on the fragile quantum system

00:59:55.500 --> 00:59:57.000
to hold the memory over long periods.

00:59:57.220 --> 00:59:58.860
- Because of the coherence time limits we discussed.

00:59:59.140 --> 01:00:01.240
- Right, so you feed the data to the quantum chip,

01:00:01.380 --> 01:00:03.200
let it scatter into 1024 dimensions,

01:00:03.620 --> 01:00:04.420
and you immediately measure it.

01:00:04.760 --> 01:00:07.980
Then you hand those measurements off to a classical co-processor like a standard

01:00:08.260 --> 01:00:08.680
silicon chip.

01:00:09.140 --> 01:00:13.840
Exactly. The classical machine takes over, handles the temporal memory fading

01:00:14.120 --> 01:00:17.960
natively in software, and calculates the final linear regression mapping.

01:00:18.250 --> 01:00:22.500
That makes total sense. It's like having a brilliant, highly specialized artist who

01:00:22.500 --> 01:00:27.500
can paint incredibly complex canvases in a millisecond but has absolutely no memory.

01:00:28.200 --> 01:00:32.960
So you pair them with a very organized archivist to remember everything and catalogs the paintings.

01:00:33.360 --> 01:00:34.020
That's a great analogy.

01:00:34.720 --> 01:00:38.200
You use the QANO chip as an accelerator just for the messy non-linear lifting,

01:00:38.720 --> 01:00:40.340
and the silicon chip handles the timeline.

01:00:41.080 --> 01:00:44.860
It completely bypasses the worst of the quantum coherence time limitations,

01:00:45.440 --> 01:00:48.080
while still leveraging the immense power of the Hilbert space.

01:00:48.540 --> 01:00:54.480
And this hybrid approach is what companies like Quera and IBM are focusing on for near-term, real-world utilities.

01:00:54.800 --> 01:00:56.540
Yes, it's the most practical path forward.

01:00:56.620 --> 01:00:56.820
Okay.

01:00:57.619 --> 01:01:01.099
I want to take a step back from the hardware for a moment as we wrap up today because as

01:01:01.119 --> 01:01:06.359
we look at all these wildly different physical platforms from vibrating atoms to pulses of

01:01:06.440 --> 01:01:11.420
light to magnetic liquids, the implications of physical reservoir computing start to really

01:01:11.660 --> 01:01:14.520
blur the lines between what is technology and what is biology.

01:01:15.079 --> 01:01:20.579
It forces a profound, almost philosophical shift in how we view the very nature of computation

01:01:20.960 --> 01:01:21.180
itself.

01:01:22.500 --> 01:01:26.380
For the last 70 years since the invention of the transistor, we have viewed computation

01:01:26.619 --> 01:01:31.240
as something strictly synthetic, something a human invented etched onto a silicon wafer,

01:01:32.160 --> 01:01:35.220
executing rigid binary algorithmic logic gates.

01:01:35.799 --> 01:01:39.299
True or false, one or zero, an entirely artificial construct.

01:01:39.740 --> 01:01:39.960
Exactly.

01:01:40.400 --> 01:01:44.180
But physical reservoir computing, and QRC in particular,

01:01:44.619 --> 01:01:47.140
proves that computation is not just a human invention.

01:01:47.880 --> 01:01:50.760
It is a natural, inherent property of physical systems.

01:01:51.359 --> 01:01:52.819
The universe computes constantly.

01:01:52.960 --> 01:01:56.319
A physical bucket of water responding to a stone falling into it

01:01:56.460 --> 01:01:57.960
is actively processing information.

01:01:58.220 --> 01:02:01.920
A network of entangled quantum spins reacting to electromagnetic pulse

01:02:02.380 --> 01:02:03.960
is actively processing information.

01:02:04.600 --> 01:02:09.020
We didn't invent the processing, we're just finally learning how to build sensors to read the output.

01:02:09.190 --> 01:02:12.680
And if we draw a direct parallel between these physical reservoirs and the human brain,

01:02:13.190 --> 01:02:16.460
the brain itself isn't a rigid digital computer running binary code.

01:02:16.490 --> 01:02:18.560
It doesn't have a CPU and a separate RAM model.

01:02:25.080 --> 01:02:28.260
The brain itself isn't a rigid digital computer running binary code.

01:02:28.260 --> 01:02:30.440
It doesn't have a CPU and a separate RAM module.

01:02:36.980 --> 01:02:41.280
No, the brain is the ultimate natively dynamic physical system.

01:02:42.380 --> 01:02:46.960
Neuro science suggests that the human brain exists precisely at that edge of chaos we discussed

01:02:47.220 --> 01:02:47.339
earlier.

01:02:47.680 --> 01:02:47.760
Right.

01:02:48.000 --> 01:02:53.640
It harnesses natural physical nonlinear dynamics, relying on the fading memory of billions of

01:02:53.680 --> 01:02:59.020
organic neurons and synapses to process the continuous, chaotic time series data that makes

01:02:59.020 --> 01:03:00.920
up our entire sensory experience.

01:03:01.400 --> 01:03:06.880
So when we use a physical quantum reservoir to decode a human EEG or an optical reservoir

01:03:07.099 --> 01:03:08.320
to track heart arrhythmia,

01:03:08.900 --> 01:03:11.760
We aren't using an artificial computer to analyze biology.

01:03:12.480 --> 01:03:17.240
We are essentially using one naturally occurring complex physical system to understand another.

01:03:17.520 --> 01:03:22.020
We are matching the inherent complexity of biology with the inherent complexity of quantum physics,

01:03:22.260 --> 01:03:26.260
entirely bypassing the artificial bottleneck of classical digital logic.

01:03:26.600 --> 01:03:28.240
is an incredibly elegant symmetry.

01:03:28.900 --> 01:03:31.340
It has been an absolutely wild journey today.

01:03:31.660 --> 01:03:36.540
We started with the simple, intuitive idea of dropping a stone into a pond just to watch the ripples spread.

01:03:37.040 --> 01:03:41.460
We scaled that concept up to tracking the entangled, wave-interfering physics of seven atoms

01:03:41.780 --> 01:03:44.020
standing a mathematical space of over a thousand dimensions.

01:03:44.900 --> 01:03:50.320
We navigated the sheer unyielding chaos of biological noise from penicillin fermentation

01:03:50.780 --> 01:03:52.460
to clinical trial failures.

01:03:52.710 --> 01:03:56.780
And we've arrived at a near future where the chaotic rhythms of the human heart might

01:03:56.780 --> 01:04:02.120
be decoded instantly by the natural, tunable vibrations of a quantum circuit on your wrist.

01:04:02.390 --> 01:04:07.040
The convergence of fundamental physics, high-dimensional information theory, and clinical

01:04:07.220 --> 01:04:08.580
medicine is happening right now.

01:04:09.000 --> 01:04:12.780
And reservoir computing is providing the architectural bridge to make it possible.

01:04:13.050 --> 01:04:13.640
It really is.

01:04:14.030 --> 01:04:18.540
And so as we close out this deep dive, we want to leave you, the listener, with a final,

01:04:19.350 --> 01:04:23.780
lingering thought to mull over as you go about your day in the lab or the clinic or just out

01:04:23.880 --> 01:04:24.220
in the world.

01:04:25.250 --> 01:04:28.480
We have seen today that computation is not confined to human-made silicon.

01:04:29.200 --> 01:04:33.720
If a bucket of water, a slice of optical glass, a vial of magnetic liquid, or a handful of

01:04:33.800 --> 01:04:37.560
quantum spins can inherently process information, learn from the past, and remember.

01:04:38.480 --> 01:04:41.200
is intelligence something we are painstakingly building from scratch?

01:04:41.780 --> 01:04:44.520
Or is it simply a fundamental universal property of matter

01:04:44.750 --> 01:04:46.780
that we are just finally learning how to tap into?

01:04:47.640 --> 01:04:50.359
No, that's a question that is going to stay with me

01:04:50.410 --> 01:04:52.040
and I think it should stay with you too.

01:04:52.640 --> 01:04:55.060
Not as a cosy, philosophical course card or what,

01:04:55.540 --> 01:04:57.079
but as a real research prompt.

01:04:57.680 --> 01:05:00.880
We've seen in this series the boundary between doing computation

01:05:01.109 --> 01:05:03.800
and just being physical stuff is much blurrier

01:05:04.400 --> 01:05:05.579
than our textbooks suggest.

01:05:06.619 --> 01:05:08.180
Also along the way I think we do

01:05:08.200 --> 01:05:11.980
It's covered that tools like the plexity and notebook lm are actually useful companions

01:05:12.210 --> 01:05:13.780
in this space, not as oracles.

01:05:15.210 --> 01:05:19.680
But as ways to hold, let's say, a lot of moving parts and a working memory while we

01:05:19.780 --> 01:05:19.880
think.

01:05:20.520 --> 01:05:24.740
If you want to go deeper, the six part sub-stack series from Classic Lacos to Quantum Dynamics

01:05:24.950 --> 01:05:27.099
is waiting for you and the link is in the show notes.

01:05:27.540 --> 01:05:30.839
I'm Thomas Aema, thanks for staying with me and being weird.