- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
This paper investigates how much of a convolutional layer’s own input can be reconstructed from its output. The proposed approach samples local input patches, encodes them using the layer’s transformation, and trains a decoder to reconstruct the patches. Recoverability is measured using held-out . The authors compare pretrained, randomly initialized, and reconstruction-optimized encoders, and introduce random controls that match pretrained singular vectors, singular-value spectra, activation sparsity, or BatchNorm parameters. Experiments cover nine ImageNet-pretrained CNN architectures and additional VGG16 and ResNet18 variants trained on CIFAR-10 and CIFAR-100. The reported findings are that recoverability generally decreases with depth, effective compression is a more consistent correlate of recoverability, and the pretrained-minus-random gap varies across architectures. Property-matching controls exhibit more consistent directions of change. The authors explicitly distinguish their decoder-relative reconstruction metric from Shannon information and investigate this distinction in the appendix.
Critical strengths
-
The local reconstruction target is well motivated and useful. Reconstructing a layer’s own input, rather than the original image from a deep representation, reduces the confounding effects of preceding transformations. Comparing encoders on identical input patches also makes the comparisons more interpretable. The probe size is independent of the feature map’s spatial dimensions, enabling evaluation across a reasonably broad set of CNN architectures.
-
The paper includes substantial controls and explicitly investigates its metric’s limitations. Beyond pretrained-versus-random comparisons, property-matching controls offer a useful way to investigate the sources of recoverability differences. The appendices examine patch sampling, decoder capacity, and alternative random initialization. They also demonstrate that recoverability and mutual information can rank encoders differently. This explicit acknowledgment makes the scope of the contribution more precise.
Critical weaknesses
-
Reported effective-compression values conflict with the stated definition and architectural dimensionality. Section 4.2 defines effective compression as , where is the number of principal components explaining 98% of input-patch variance. For a convolution, this necessarily implies $ k_{98}\leq C_{\mathrm{in}}, \frac{k_{98}}{C_{\mathrm{out}}} \leq \frac{C_{\mathrm{in}}}{C_{\mathrm{out}}}. $ However, page 7, lines 326-328, reports effective-compression values of 0.88–0.94 for expansion layers in architectures including ResNet50. In standard torchvision ResNet50, the bottleneck expansion convolution has , so the stated definition gives an upper bound of 0.25. This could be a textual error, or it could reflect a discrepancy in the PCA input, denominator, or layer identification. The manuscript does not resolve these possibilities. Since effective compression is a central explanatory variable, its calculation and the associated results need verification before I can accept this conclusion.
-
The substantially positive held-out under shuffled targets remains unexplained. Appendix A.2 reports mean held-out shuffled-target values of 0.107 and 0.124 for the and decoders, respectively, with some values above 0.2 in Figure 8. If shuffling removes the input–target relationship, training and validation are independent, and uses a coordinate-wise mean predictor as its zero baseline, ordinary training-set memorization does not explain this validation performance. Section 3.3 specifies without stating its reduction axes. If variance is computed around a single global scalar mean, differences between coordinate means can allow a constant vector predictor that ignores the representation to achieve positive scores. Furthermore, the residual decoder is initialized from a fitted linear solution, and it is unclear whether this solution is recomputed using shuffled targets for the control. Consequently, the observed behavior cannot yet be interpreted simply as “memorization.” The authors should distinguish normalization, shuffling, and initialization effects. This matters for absolute recoverability and cross-layer comparisons, beyond the appendix diagnostic itself.
Questions for authors
-
Could you provide the layer name, , , , and PCA input-matrix shape for a representative ResNet50 expansion layer discussed on page 7? How are the reported values of 0.88-0.94 compatible with the dimensionality bound above? If this reflects a calculation or layer-matching issue, do Figure 3 and the main correlation results hold after correction?
-
What are the exact reduction axes for MSE and variance? What is shuffled, when is it shuffled, and which targets are used for validation? Is the linear initialization refitted on shuffled training targets? For representative layers with high shuffled-target scores, could you report a constant mean-vector baseline and results using $ R^2 = 1-\frac{\sum_i|x_i-\hat{x}_i|^2} {\sum_i|x_i-\bar{x}|^2}, $ and establish whether the principal comparisons remain unchanged?
-
The introduction does not establish the significance of the chosen measurement sharply enough. The introduction moves from biological information bottlenecks to mutual information, representation inversion, and probing, but provides a less developed explanation of why recovering a layer’s own input patch is scientifically useful. The paper should identify the precise question this local measurement can answer and explain what new understanding it provides beyond dimensionality reduction and reconstruction error. The biological motivation could be shortened to make room for this argument.
-
The method description is insufficiently explicit about how the probe corresponds to different architectures. Section 3.2 describes a fully connected encoder optionally followed by BatchNorm and an activation, and asserts equivalence to the convolutional layer’s output. However, the evaluation includes depthwise convolutions, residual architectures, and networks with different normalization and activation placement. A compact implementation table should specify the input and output endpoints, handling of grouped/depthwise connectivity, BatchNorm mode and statistics, and whether residual addition is included. These details are necessary to assess the claimed equivalence.
-
The depth plots omit the coordinates needed to identify the reported effects. Figures 2, 5, 13, 14, and 15 show progression over depth without numerical layer ticks or sufficiently identifiable stage labels. Readers can see trends but cannot reliably associate spikes, drops, or transitions with particular layers. The figures should include layer indices or stage boundaries and identify important exceptions discussed in the text. For architectures with parallel branches, the authors should also explain how the displayed ordering is constructed: connecting every point with a line can otherwise suggest a sequential processing path that the architecture does not have.