Layer-wise Input Recoverability in Deep Neural Networks
This study investigates how much information about a convolutional layer's input can be recovered from its output. Researchers measure this 'layer-wise input recoverability' by seeing what fraction of the input patch variance a decoder can reconstruct from the layer's output. Findings show recoverability generally decreases with network depth and is strongly correlated with the layer's compression ratio, indicating that effective compression limits information retention.