Data repository for "Loss Behavior in Supervised Learning With Entangled States"

<p>
Replication code and experiment result data for training Parameterized Quantum Circuits (PQCs) with entangled data. The experiments evaluate the structure of the loss landscape during training based on the training sample that is used for training.
</p>

<p>
The combined experiment and data extraction scripts are contained in
<code>experiments_and_data_extraction.zip</code>. The experiment results that are used in the related publication are contained in <code>raw_results.zip</code>.
</p>
<p>
Experiment/Reproduction setup (<code>experiments_and_data_extraction.zip</code>):
<ul>
<li><code>env.sh</code>: Contains specification of experiment environment. Adjust the number of available CPUs as needed.</li>
<li><code>run_exp_5.py</code>: Entrypoint for experiments for training 5-qubit PQCs. The used PQCs and hyperparameters can be adjusted in this file.</li>
<li><code>data_extraction.py</code>: Extraction and aggregation of experiment results in preparation for the contained plots. Note that this step was already executed on the results in <code>raw_results.zip</code> and can be omitted if the plots should be reconstructed using the available data.</li>
<li><code>states_3_qubits.py</code> and <code>states_5.npy</code> contain input states used for the experiments (sorted by number of qubits).</li>
</ul>
</p>
<p>
Plots:
<ul>
<li><code>improvement_neighborhood.pdf</code>: Comparison of the improvement in a neighborhood ordered by the Schmidt rank of the input. Created by <code>improvement_plot.py</code>.</li>
<li><code>improvement_by_schmidt_rank.pdf</code> and <code>improvement_by_entanglement_entropy.pdf</code>: Comparison of improvement order by entanglement measures. Created by <code>improvement_by_entanglement.py</code>.</li>
<li><code>expressibilities_separate.pdf</code>: Comparison of expressivity by PQCs. Created by <code>expressibility_plots.py</code>.</li>
<li><code>distance_to_zero_comparison.pdf</code>: Comparison of distance to minimum in loss landscape. Created by <code>distance_to_zero_plot.py</code>.</li>
</ul>
</p>