A unifying framework from neural superposition to sparse interpretable codes
et al. present a unifying framework for superposition in neural networks. Their three-step approach clarifies how latent features can be identified, disentangled and assessed.
