MCPC with and without PC warm-up inference steps trained on the MNIST dataset.

<p>We train an MCPC model on the MNIST dataset with and without warm-up steps. Moreover, we evaluate a range of inference step counts. When the model is trained without warm-up steps, the inference process includes MCPC mixing steps followed by a single MCPC sampling step. However, when the model undergoes training with PC warm-up steps, the initial half of the inference steps consist of PC inference steps, while the remaining inference steps are MCPC mixing steps and one MCPC sampling step. The model parameters for training can be found in <a href="http://www.ploscompbiol.org/article/info:doi/10.1371/journal.pcbi.1012532#pcbi.1012532.s004" target="_blank">S4 Appendix</a> and correspond to the parameters that maximise the FID measure. This figure demonstrates that using PC warm-up inference steps results in improved performance with a limited number of total inference steps, diminished performance with 100 or 200 inference steps, and comparable performance with a large number of inference steps. Ultimately, this result shows that warm-up steps are not always beneficial and should be considered for each learning task separately. The results are shown for three initialisation seeds. The dots show the mean result while the error bars show the standard deviation.</p> <p>(PDF)</p>