This repository contributes to the development of a derivative-free neural network optimiser, using the parallel tempering Markov Chain Monte Carlo (MCMC) algorithm implemented in the MOSAICS software (http://www.cs.ox.ac.uk/mosaics/).
We compare the performance of a standard gradient-based optimiser to the proposed one by scrutinising the trajectory in a 2D projection of the feature space (visualise_trajectory.py) as well as network performance and memory cost.