How do I use multiprocessing to speed up training of a neural network

Question

I created a simple MLP neural network to create a model for the MNIST dataset, but I found it takes a while (~30 min) to loop through all 60k images, but that is using only one process. I want to use multiprocessing to speed up training by using multiple processes.

import numpy as np
import multiprocessing as mp

num_images = 60000

weights = np.random.rand(3)

x = np.array([0.01, 0.01, 0.01])

images = [i for i in range(num_images)]

def train(images):
    global weights
    for i in images
        weights = weights + i * x

This is a simple example because I wanted to simplify to figure out the problem, but I just want to use multiprocessing to split up the images (represented by integers for simplification) into equal-sized smaller lists and each process to apply the train function simultaneously, but they also have to be able to change the global variable weights.

I can't use multiprocessing.Value() because it requires a simple object format like int or float, and I can't use numpy arrays.