Projecting into a bigger, cleaner space
If a network is packing more features than it has neurons into overlapping directions, one way to fight back is to project those activations into a much larger space where there's finally enough room for each feature to get something closer to its own dedicated direction. This is the idea behind a sparse autoencoder, a small auxiliary model trained not on the original task, but on the internal activations of the network you're trying to interpret.
A sparse autoencoder takes a layer's activations as input, expands them into a much larger intermediate representation, and reconstructs the original activations as output. The 'sparse' part is the key constraint: the training objective pushes most of those intermediate units to be zero for any given input, so only a small number are active at once. That sparsity pressure is what encourages each unit in the expanded space to specialize, since the model is penalized for spreading activation across many units when a few would do.