An introduction to MB-QML¶

Author(s): Luis Mantilla

Quantum machine learning (QML) is a field that studies how to use parametrized quantum circuits to learn to identify patterns in quantum data. In measurement-based qunatum machine learning (MB-QML) [1], we use a MBQC circuit with parametrized measurement angles to solve QML problems.

In mentpy, MB-QML models are defined using the MBQCircuit class. We can define a model from scratch or use one of the templates provided in mentpy.templates. Here, we use the MuTA template with two input qubits, and fix two of the parameters to be fixed (qubits 3 and 8).

In [1]: import numpy as np

In [2]: import mentpy as mp

In [3]: gs = mp.templates.muta(2,1, one_column=True)

In [4]: gs[3] = mp.Ment('X')

In [5]: gs[8] = mp.Ment('X')

In [6]: ps = mp.PatternSimulator(gs)

In [7]: mp.draw(gs)
Out[7]: (<Figure size 800x300 with 1 Axes>, <Axes: >)

To optimize the parameters of the model, we need to define a loss function. Here, we will use the average infidelity between the target states and the output states of the model.

In [8]: def loss(output, target):
   ...:     avg_fidelity = 0
   ...:     for sty, out in zip(target, output):
   ...:         sty = mp.calculator.pure2density(sty)
   ...:         avg_fidelity += 1-mp.calculator.fidelity(sty, out)
   ...:     ans = (avg_fidelity/len(target))
   ...:     return ans
   ...: 

In [9]: def prediction_single_state(thetas, st):
   ...:     ps.reset(input_state=st)
   ...:     st = ps(thetas)
   ...:     return st
   ...: 

In [10]: def prediction(thetas, statesx):
   ....:     output = [prediction_single_state(thetas, st) for st in statesx]
   ....:     return output
   ....: 

In [11]: def cost(thetas, statesx, statesy):
   ....:     outputs = prediction(thetas, statesx)
   ....:     return loss(outputs, statesy)
   ....: 

Be aware that the loss function is a global operation, which can induce barren plateaus. However, we will ignore this issue for now. Having defined a model and a loss function, we can now use some data to train our model. We will use the generate_random_dataset() function to generate a random dataset of states \(\left\{(\rho_i, \sigma_i)_i \right\}_i^{N}\) where the input and target states are related by a given unitary \(\sigma_i = U \rho_i U^\dagger\).

runs_train = {}
runs_test = {}

NUM_STEPS = 100
NUM_RUNS = 10

for i in range(NUM_RUNS):
    random_gate = np.kron(mp.utils.random_special_unitary(1), np.eye(2))
    (x_train, y_train), (x_test, y_test) = mp.utils.generate_random_dataset(random_gate, 10, test_size = 0.3)

    cost_train, cost_test = [], []

    def callback(params, iter):
        cost_train.append(cost(params, x_train, y_train))
        cost_test.append(cost(params, x_test, y_test))

    theta = np.random.rand(len(gs.trainable_nodes))
    opt = mp.optimizers.AdamOptimizer(step_size=0.08)
    theta = opt.optimize(lambda params: cost(params, x_train, y_train), theta, num_iters=NUM_STEPS, callback=callback)
    post_cost = cost(theta, x_test, y_test)

    runs_train[i] = cost_train
    runs_test[i] = cost_test

Finally, we can average over the runs and plot the results!

An introduction to MB-QML¶

References¶