Transcript
This transcript was autogenerated. To make changes, submit a PR.
I wanna talk about hybrid quantum, classical neural networks and
their application to the field of Cy Hybrid Engineering.
As you'll know, like the quantum computing field is picking up,
Google has launched a new chip Willow and so has Amazon recently,
and they have passed the benchmarks.
But to make use of them practically now for especially the emergence of AI
and machine learning, we are looking into the hybrid approach of hybrid
quantum and classical neural networks.
I have briefly split the agenda into hybrid quantum classical neural networks.
I'll say HQC and Ns.
Very long to pronounce the challenges that are in this
approach, the practical applications.
And the SRA aspect of it.
Okay, so
the left column talks about the quantum advantages.
So quantum computing takes advantage of the quantum principles of physics.
Which is super position.
Super position has the ability of data to be in all the possible states
at once before it's being observed.
If you take in binary, we call it a zero or one, it could be a zero or one, or.
Both zero and one before it is being observed.
So this arises to the concept of parallel invoices and the ability to solve the
problem parallel, in parallel invoices.
And quantum ment is the ability of one qubit.
Qubit is the basic unit of like a bit of quantum computers.
Ability of one qubit to to affect the other, to the same
state respect of their distance.
This adds tremendous power to quantum computing.
And the modeling, the modeling of quantum states using hilbert spaces, provides
unprecedented power to solve problems.
And as the classical strengths, it's well established like what charge
GPD is using or, any other tropic or any other your network is using.
They're using regular binary com, or not regular, but established
high GPU classical computers when you fuse both of them.
So basically, you are using the accommodation of classical.
You basically provide the problem in classical, which
is encoded into quantum state.
And you return the result back in classical after validation
that provides unprecedented problem solving capabilities.
Okay.
Now enters the.
Parameterized quantum circuits, but they are quantum algorithms that incorporate
tunable parameters within quantum gates, enabling optimization of tasks.
As you can see in this dis, in this graphic, the presentation, so there is a
feature mapping part below which basically maps the classical part features to the,
classical features to the quantum states and the quantum parameters are tunable
elements with within the quantum circuits that enable optimization for specific
tasks such as quantum machine learning, classical gradient based techniques.
The one that is used in the last piece classical optimization, are tuneable
elements within quantum circuits that enable optimization for specific
tasks and techniques like gradient doesn't are foundational, which
it iteratively are just parameters to minimize the cost function.
Okay.
I think the slide is slightly.
Your slide is slightly bigger.
So this one, it this slide compares the HQ CNNs with the traditional
classical neural networks.
And you can see that HQ CNNs consistently outperform the.
The classical computers and their, the Neur network approach, despite
having noisy data, and we still have stability issues with qubits, which
just had to be resolved, especially in a complex, high dimensional tasks like
medical diagnostics or drug discovery.
So it's got a huge implication on basically us think of it like this,
if you can find the vaccine for Q Covid within hours rather than months.
See the impact can have on the number of lives that we could save.
So there is a potential of this approach.
Okay, I'm going to reshare.
Okay,
so now comes how do we train
quantum circuit, steep quantum circuits.
First is the circuit initialization.
Where you structure quantum layers with appropriate gate operations
and parameter optimization, applying grade dissent based methods.
Next is error mitigation.
And finally, you evaluate the model.
Basically, you take a known known benchmark and you, you run it both
in the classical as well as on the hybrid quantum computer, and then
confirm that the quantum works.
Or it is within the benchmark of the classical, computers for accuracy.
Okay?
Yeah.
So the advantages of high dimensional computation, as in the sa in the slide.
It speeds up like qu the things that we talked about.
It's made up of qubits and it has got the the typical quantum
advantages of superposition and entanglement and quantum parallelism.
So basically it is almost like having parallel invoices to solve a problem.
You can think about the consequence of that thought and, especially
on high dimensional data.
High dimensional data which requires basically a lot of possible
ways to look into the solution.
They offer, they're very greatly efficient, especially compared
to the classical representation, efficient optimization, quantum
gardens boss solution landscapes, like assume you have you want to come
up with a financial modeling and.
You want to say what happens when I do it on this day?
If you want to take the 365 days of a year and calculate the interest and the
subsequent pension that you'll receive after six 60 years, the classical computer
will struggle, but whereas the quantum can easily model that in just a fee.
Yeah.
The dimensionality advantage hybrid architectures seamlessly integrate
both quantum and classical data sets.
Basically it maintains the regular computing advantages, the practical use
of it, and also the quantum advantages, which is the ability of IT to solve
problems with high emission spaces.
Yeah.
Now let's look at some of the applications.
It's very exciting.
Like we, whatever we talk till now, it's more like a theory,
but this is practical, right?
Especially in medical diagnostics, like if to read an MRI scan or a or
combining multiple data, multiple scans together and deriving conclusions
out of it, or, which can be critical for the patient's diagnosis.
So this offers unparalleled potential drug discovery, as I mentioned.
Like we.
Already we are doing a little bit like Covid was done pretty fast, but quantum
can do it within no time, almost like minutes or day or hours, which basically
gives us the ability to think about novel drugs, which can completely change
the life of people and cure diseases.
Yeah.
The other frontier is the personalized medicine.
You have multiple data sets, right?
You have the clinical data where you go to a GP or have your medical test and
the environment that you live in your pre predisposition, your genomic data.
The HQ n they can combine all of these and produce very personalized.
Medicine, like the one that perfectly works for you, depending
on all the parameters, which may not be so easy to do right now.
As I talked the model the potential applications are in financial modeling.
Gimme one second.
Trying to move this.
Okay.
Financial modeling.
Then again the supply chain logistics it's like an operational such problem.
You have limited resources and what is the best way to, for the cost being
optimized or the faster delivery.
For critical medical emergencies or it's a very good area to
work on supply chain logistics.
And it can help itself.
It can design efficient quantum and classical circuit layouts, which
can make it even more efficient.
Efficient.
And also one benchmarks like they have demonstrated that for 200.
Plus training iterations.
The variance is only between 2%, which is pretty high for accuracy.
Stability,
okay.
The way we train a quantum hq, hybrid QU hq, CNNs is.
We have to first have the dataset create a classical dataset, and then
we have to prepare it into a format that's digestible by the quantum
computers, like amplitude amplitude.
So basically you convert that into a voltage or a, or direction, like
a phase sample, and then encoding the feature ma maps to preserve
the critical relationships.
Okay.
And next, like executed forward pass is, they do one iteration, move forward
to perform complex transformations on the encoded data, and then the data is
sent back to classical back propagation.
Propagate propagation is enabling efficient organization across the quantum
and class, so quantum has computed the data is available in the quantum format.
You need to pass it on to the classical format and hybrid optimization.
Systematically refined parameters to iterative quantum classical feedback log.
So basically the earlier ones that we created, indecent, we saw that, right?
So you keep on doing iteratively to optimize the hybrid system.
It's not very straightforward.
There are a lot of challenges.
It's in the much, still in the state of infancy, but it's quite an interesting
problem and eventually we'll get there.
So one is quantum de coherence, like the challenges I mean despite we say
that the qubits have we have made it much better than what it was, but
still they suffer from instability.
And the error mitigation techniques are still work in progress.
And
scalability is one issue.
And barren plat like when the.
Is a critical training problem in quantum machine learning, QML, specifically in
training p QCs, parameterize quantum circuits using gradient based methods.
So basically when the when the quantum circuit becomes deeper and deeper the.
The gradient vanishes.
So that is a problem, that's a common problem or a problem
that's studied deeply.
Barr plats and the classical quantum interface, the data transfer between
our classical systems and quantum.
They create bottlenecks and hybrid algorithms must minimize the server.
Okay, so the future outlook, as Google.
Amazon, Microsoft, everyone is working towards, even in media, they're working
towards more quantum computing chips with more qubits, coherence and gate fidelity.
And on the software side, the algorithms are also improving.
No hybrid architectures are created.
And the other thing is know we are talking this is now happening in
labs in some universities or in some, like top top companies, labs, or
RD divisions, but it has to become a widespread industry adoption.
And the quantum and machine learning has to become a standard, HQs,
but they will surely revolutionize the computational approach.
And when this matures solving the problems that we could not solve earlier
or could not even think about earlier.
Now the relevance of, now we, we discussed about quantum, we discussed about
classical and how HQ c, hybrid quantum quantum computing neural network model
like computers how how they work and how they can help us in drug discovery and
personalized medicine, financial modeling.
We talked about all that.
Now.
What are the problems from Ari?
Perspective?
One is the handoffs is a point of failure and because the movement
of data from classical to quantum, so that's a point of failure.
So we from ARI perspective systems have to be Ari, have to design robust
interfaces and ensure that, that there is no data loss or data transformation.
The other, one other important piece I want to talk about is the observability.
Okay?
Now, if you have a cloud computing system, you have a way to observe it.
So how do you observe a quantum state?
How do you observe quantum noise?
How do you observe quantum decoherence?
And how do you observe latency?
How do you know the models performing well?
Is it not drifting away?
How do you know all those things?
So observability is a many important part.
Part of it.
And error mitigation strategies, right?
If there is an error during a handoff or initial data encoding, how do you come
up with ways to, mitigate the errors?
Okay.
And scalability is another problem.
And deploying.
Okay.
There suppose be there is a new model that's getting deployed, so how do
you deploy it into the quantum space?
Do we need any special orchestration techniques, ization strategies
where the have to come up with new models and something that they
have not thought about till now.
And
SREs were supporting that they have to.
They have to ensure that they can deliver a level of repeatability, load drift.
And even when the system is loaded with a heavy model or load tested
parallelly, it should not drift.
And it, it produces the same consistent, pro consistent
performance even during load.
Okay SRE must manage the handoffs, error mitigation as we talked
about earlier, and monitoring and observability and scalability, right?
Like how do you suppose you want more clusters, how do you scale on demand?
Can you do it within few milliseconds so that there is no loss in the performance.
Those kind of things are very critical for for an SRE and I put together, put
put together a set of references for you, the one from IPM and the parameterized
quantum circuits, and, or how to build reliability by IBM and practical use.
Recently U Austin, they came up with a way to test whether the numbers that
are generated by a quantum computer is really random using a classical computer.
I put across some references or you like it and it's, you
learn something out of it.
Yeah, that's all I have.
Hope you like it.
Thank you very much.