1 of 10

Swagat Kumar

PhD Student�swagat.kumar@insait.ai

Est. 2022

RhoDARTS: Density Matrix Simulations to Facilitate Quantum Architecture Search

2 of 10

The Team

Jan-Nico Zaech

Postdoc Researcher

Luc Van Gool

Professor (Computer Vision)

Colin M. Wilmott

Senior Lecturer (Mathematics)

Swagat Kumar

PhD Student

3 of 10

Neural Network Architecture Search

Which architecture (choice of operations) gives the best performance?

Dense Layer

Convolutional Layer

Attention Layer

Dense Layer

Convolutional Layer

Attention Layer

Input

Hidden Layer

Output

4 of 10

Differentiable Architecture Search (DARTS)

Input

Hidden Layer

Output

Each hidden layer is a sum of the outputs of every operation, weighted by the probabilities to select them.
The loss function is used to train the architecture and network parameters simultaneously.

5 of 10

Quantum Architecture Search (QAS)

?

Gate Pool

What architecture (choice of gates) produces the desired state?

6 of 10

DARTS does not work with state vectors

Not a valid quantum state!

?

7 of 10

Differentiable QAS

?

Search Space

?

Sampled Circuits

Approximate architecture gradients by evaluating the sampled circuits.

Search space grows exponentially with the number of qubits and circuit depth.
Accuracy and rate of convergence depends on the number of samples taken in each step.

8 of 10

Density Matrices

Model classical uncertainty about quantum states.
Useful in modelling noisy quantum systems.

Valid quantum state!

?

9 of 10

ρDARTS

arXiv link