Here, let's plot the weights of the final layer with respect to the output (reconstruction input data). In the current scenario, 900 is the number of nodes in the hidden layer and 784 is the number of nodes in the output (reconstructed) layer.

In the following image, the first 400 nodes in the hidden layer are seen:

Here, each tile represents a vector of connections formed between a hidden node and all the visible layer nodes. In each tile, the black region represents negative weights (weight < 0), the white region represents positive weights (weight > 1), and the grey region represents no connection (weight = 0). The higher the positive value, the greater the chance of activation in hidden nodes, and vice versa. These activations help determine which part of the input image is being determined by a given hidden node.