- Rating
- 1
- Confidence
- 2
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The authors evaluate intermediate layers of DNNs to find stable representations. They define stability as representations that give similar results across different instances of the same network trained for the same task. They analyze six ImageNet-trained architectures with five instances each. They compare layers across instances with two similarity measures (JSD with BayesCompare, and linear Centered Kernel Alignment) on MS COCO images. From this analysis, they empirically derive three criteria that characterize the most stable layers. Additionally, they modify ConvNeXt Tiny based on these criteria, and a stable layer type emerges, as predicted. Finally, they show on two downstream tasks (fMRI brain encoding and transfer learning) that stable layers give more consistent results across instances and comparable or better performance than non-stable layers.
Critical strengths
The paper tests stability in two very different applications: fMRI brain encoding and transfer learning on CUB-200-2011. In Table 1, stable layers have a higher mean accuracy in most of the cases. However, this strength is limited. The results depend on the stable-layer selection; many differences are small, and there are clear exceptions.
Critical weaknesses
-
The evaluation uses only five instances per architecture: four trained by the authors and one pretrained PyTorch model. The four trained instances differ only by random initialization and “minor modifications in their base training recipes” (lines 106–107). Their accuracies are within about one percentage point. The modified ConvNeXt is tested on only four instances. The current results show stability across small training differences. However, it is unclear whether this stability holds across larger training differences. More instances with clearly different training recipes would make the conclusions stronger.
-
It is unclear how the stable layers are selected. Table 3 orders the layers by the average of JSD and CKA, but the paper gives no rule for which layers are stable. The number of stable layers varies from zero to three. With this average, ResNet50 matches the paper’s labels only with the top one, ViT only with the top two, and EfficientNet only with the top three. In Swin-T and Swin V2, the rule is even more unclear, because Linear1 has the lowest average of all block layers but is marked as non-stable. In the modified ConvNeXt, LayerNorm1 (JSD 0.1426, CKA 0.0849) is marked as stable. However, GeLU1 has almost the same values (JSD 0.1362, CKA 0.0906) in the original ConvNeXt and is marked as non-stable. This is critical because the three criteria, the ConvNeXt test, and the downstream results all depend on these labels.
Questions for authors
-
Do the findings still hold if the training recipes change significantly?
-
Could the authors state the rule they used to classify layers as stable? Why is Linear1 not selected in Swin-T and Swin V2, although it has the lowest average self-distance (JSD and CKA) in both models?