Deep neural networks remain hard to interpret. A new arXiv preprint, "Layer-wise Input Recoverability in Deep Neural Networks," takes up this problem directly, noting in its abstract that the inner workings of deep networks are still not well understood.
The paper situates itself relative to the information bottleneck framework, which characterizes deep networks by the mutual information between their hidden representations and the network's input or output. The proposed approach appears to shift the focus to input recoverability: rather than asking only how much information a layer retains, it asks how much of the original input can be reconstructed from that layer.
The source abstract is truncated, so specific methods and results are not available. Based on the title and opening, the contribution is a layer-wise recoverability perspective that could complement or challenge the information bottleneck account. No other sources were provided, so there is no independent comparison to report.