v0lta / Quantum-init

Quantum random numbers and quantum initialization for convolutional neural networks using PyTorch and Qiskit.

Quantum initialization for Convolutional Neural Networks

This repository compares quantum initialization and pseudo-random initialization for a convolutional neural network on the MNIST digit recognition problem. The idea is to evaluate the quality of the pseudo-random numbers commonly used to set up neural networks before training. Quantum computers can produce perfect random numbers by measuring Hadamard-transformed QBits. This an important difference from GPU or CPU randomness, which is typically generated using some form of the Mersenne-Twister (https://en.wikipedia.org/wiki/Mersenne_Twister). The get_quantum_uniform function from qrandom.py implements a quantum random number generator. In the academic literature, this idea has previously appeared in https://link.springer.com/article/10.1007/s00500-019-04450-0 . The code in this repository repeats a variant of their CNN experiment.

Dependencies

The code in the repository depends on Qiskit and PyTorch. To install these dependencies on the command line type

   $ pip install qiskit
   $ pip install torch torchvision 

Setting up IBMQ

Head over to https://quantum-computing.ibm.com/ and set up an account. An account token is required to run this code. Take a look at https://quantum-computing.ibm.com/docs/manage/account/ to learn where to find your access key. When you have the token open our python interpreter and run

 >>> from qiskit import IBMQ
 >>> IBMQ.save_account('insert your IBMQ access-token here.')

to set up your configuration.

Experiments

Once the environment is set up, use 'mnist.py' to run the pseudo- and quantum initialized experiments by typing

  $ python mnist.py --pseudo-init
  $ python mnist.py

to run the pseudo-random and quantum-random experiment in that order. The second experiment will take a while, depending on the length of the quantum computer queue.

Results

The plots below show the test set network loss for both experiments.

Quantum-initialization-convergence:

Pseudorandom-initialization-convergence:

Known issues:

Occasionally IBMQ may assign you a machine, which produces non-uniform numbers. You can check your latest jobs at (https://quantum-computing.ibm.com/) All output bits should have the same probability. Otherwise, network convergence will likely be negatively affected. See the plot below for such an example: