Similarity of Processing Steps in Vision Model Representations

TL;DR

The paper asks whether universal representations are accompanied by universal processing steps in vision systems. It analyzes layer-wise evolution of representations across iGPT, DINOv2, ViT, and ConvNeXt using the information imbalance measure $ \Delta(A \rightarrow B) \approx \frac{2}{N}\langle r^B \mid r^A = 1 \rangle$ and neighborhood-based semantics. Key findings show a shared 'distance rule'—early layers resemble early layers of other models, middle resemble middle, and late resemble late—with notable exceptions: iGPT remains dominated by low-level organization, while DINOv2 preserves both low-level and semantic cues and classification-trained architectures shed low-level information. These results clarify which processing steps are universal, inform interpretability and architecture design, and highlight the need to distinguish between representational and processing convergence.

Abstract

Recent literature suggests that the bigger the model, the more likely it is to converge to similar, ``universal'' representations, despite different training objectives, datasets, or modalities. While this literature shows that there is an area where model representations are similar, we study here how vision models might get to those representations -- in particular, do they also converge to the same intermediate steps and operations? We therefore study the processes that lead to convergent representations in different models. First, we quantify distance between different model representations at different stages. We follow the evolution of distances between models throughout processing, identifying the processing steps which are most different between models. We find that while layers at similar positions in different models have the most similar representations, strong differences remain. Classifier models, unlike the others, will discard information about low-level image statistics in their final layers. CNN- and transformer-based models also behave differently, with transformer models applying smoother changes to representations from one layer to the next. These distinctions clarify the level and nature of convergence between model representations, and enables a more qualitative account of the underlying processes in image models.

Figures (17)

• Figure 1: Information imbalance between models at different depths; lower information imbalance for A $\rightarrow{}$ B means the neighbor structure of A is predictive of that of B. Four models are compared, in both directions. We look at the information imbalance for an early (blue circle), middle (purple square) and late layer (yellow triangle) of a first model with all layers of the second model. The title of each subgraph indicates the direction, with an arrow from the predicting model towards the predicted one.
• Figure 2: Nearest neighbors of a sporting dog (top left) image from a set of 100k images, computed using cosine similarity in the space of different layers for different models. Each column is a model, each row is a specific layer. Early layers are all second layers, and late layers are all penultimate layers to avoid effects from tokenization or detokenization on the very first and very last layers, respectively.
• Figure 3: Percentage 10 nearest neighbors of an image that share with it at least one of 9 categories of low-level features. Baseline is for the case where images are randomly spread in the space.
• Figure 4: Shared ManyNames labels divided by total labels (Jaccard Similarity) for 50 randomly sampled images and their neighborhoods. Higher Jaccard similarity means that a given neighborhood shares more labels. Shaded out regions show standard deviations across the 50 images and their 10 nearest neighbors.
• Figure 5: Density plot of classification accuracy roughness across layers (left), and examples (right) of smooth (red iGPT example) and rough (blue ConvNeXt example) probing accuracies across layers. We consider a curve "rough" if its cross-layer changes are inconsistent, which we measure with standard deviation of layer-to layer differences. In contrast, a smooth curve would consistently have the same layer-to-layer variation. The two example classes have trajectory roughness of 0.395 (pizza) and 0.080 (bloodhound), respectively, which are marked in dashed lines on the density plot.