acceptodds
Under review as a conference paper at ICLR 2027

Directional Linear Separability of Neural Representations: Geometry and Transformations

Abstract

Neural networks build representations through affine maps and nonlinear activations. Injective affine maps preserve linear separability, raising the problem of how they prepare data for nonlinear improvement and how much gain can be guaranteed before complete separation. We introduce the directional linear separability measure (D-LSM), which quantifies unavoidable competing-sample intrusion over affine halfspaces retaining every target sample, characterize its supporting geometry, and prove invariance under injective affine embeddings. For gated activations including ReLU, GELU, and SiLU, pre-activation projection bounds yield sufficient conditions for preserving all previous exclusions and recovering additional samples, with a gain bound determined by the certified recovery count. Under an aggregate-tube condition, an explicit affine construction realizes recovery with sufficient width, scaling conditions, and simultaneous multiclass guarantees through a shared layer. Exact controlled experiments compare certified and realized gains, assess certificate coverage, and exhibit bound attainment before complete separation and in affine-tube constructions. In learned Vision Transformer (ViT) representations, a feasible lower-bound estimator yields earlier post-GELU saturation certificates of exact separability, while boundary transport numerically supports affine invariance.

Primary area:
interpretability and explainable AI
Keywords:
Directional separability, representation geometry, deep representations, nonlinear activation functions, geometric certificates

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_AhTCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents the Directional Linear Separability Measure (D-LSM) to quantify competing-sample intrusion under the condition of preserving all target samples. The authors first prove that D-LSM is invariant under injective affine transformations, confirming that separability improvements come from non-linear activations. Building on this, they derive pre-activation certificates for ReLU, GELU, and SiLU, guaranteeing the preservation/increase of separation. They build a concrete "affine-tube" layer guaranteed to achieve these targeted conditions. Finally, they validate their theorems on synthetic datasets and deploy a differentiable surrogate estimator to check separability across the layers of a Vision Transformer.

Critical strengths
  • The paper manages to understand how each network operation acts on separability. They prove injective affine transformations leave D-LSM invariant. They analyze the impact of coordinatewise activations on D-LSM, deriving clear certificates.
  • The paper translates a theoretical concept (pre-activation certificates) into an implementable layer. It shows how specific weights, biases, and activations can be arranged to improve separability.
Critical weaknesses
  • While it is great to have moved from abstract theory to a concrete layer, I question whether this layer has practical utility for real-world modeling. First, the architecture's width requirement explodes with the number of intruders, making it computationally intractable for large-scale datasets. Also, the paper presents no empirical evidence that maximizing D-LSM or forcing these early saturation certificates improves classification accuracy or robustness.
  • The empirical evaluation suffers from a gap between theory and practice. First, by jumping from toy synthetic datasets directly to a Vision Transformer on Tiny ImageNet, the authors bypass canonical benchmarks (e.g., MNIST, CIFAR-10) and standard architectures (MLPs, CNNs) necessary to validate the metric. Second, the large-scale experiment relies entirely on a D-LSM surrogate estimator with no theoretical approximation bounds. Without these bounds, it remains unclear whether the ViT findings reflect true D-LSM separability or optimization artifacts of the surrogate.
Questions for authors
  • Does a higher D-LSM score or earlier saturation actually correlate with improved test accuracy? If not, what concrete utility does auditing this metric offer practitioners over standard linear probing?
  • Why were standard benchmarks (e.g., CIFAR-10) and architectures (ResNets, MLPs) omitted? Comparative audit curves are needed to ensure the observed D-LSM dynamics can be generalized.
  • Can we quantify the approximation error between your differentiable surrogate and the exact metric
Reviewer_zPbKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces a new method of linear separability and study how certain choices of activations in neural networks can be shown to preserve important properties of the separability (under certain assumptions and for certain layers). They analyze this theoretically, providing certain results about the applicability, as well as perform many experiments (both on controlled setups as well as large-scale trained ViT models), in order to show that their results hold in real settings.

Critical strengths
  1. The only critical strength I can identify in this paper is the breadth of experimental evaluation, making sure to test every part of their approach experimentally to provide verification.
Critical weaknesses
  1. The collection of assumptions and requirements for this method to hold is very restrictive, and limits its practical usage: (a) the injectivity assumption immediately prevents applying this analysis to any kind of bottleneck (even for example the second layer of transformer MLPs), and it is very common for trained networks to be low-rank (not full-rank), which means that this assumption will fail for many trained network layers. (b) the aggregate-tube condition required for Theorem 3.7 is also very restrictive. My understanding of this requirement is that it assumes that all of the data in the set RR is within some thin tube of size ϵ\epsilon larger than AA formed by hyperplanes, while also assuming that the data in AA is outside this tube. That seems very unlikely to hold in practice, and instead of demonstrating certain properties of how these layers reshape data, it instead assumes the data is already sufficiently nice, and then proves that the layers can use the data effectively. (c) Again in theorem 3.7, the requirements on κ\kappa and λ\lambda seem so large that it is unlikely that it reflects any real geometry (anything very close to the boundary, with reasonable target clearance gap, will lead to λ\lambda being in the hundreds or thousands). My understanding is that these terms will necessarily affect the size of the weight matrices, thus pushing the weights significantly larger than weights typically are in learned networks.

  2. I also have concerns about the experimental results obtained from ViT models. In a one-vs-rest setting on Tiny ImageNet, the target class only contains 500 samples, while the dimensionality of the feature space is either ~20k or ~75k. Such a small amount of samples in such a high dimensionality becomes extremely easy to separate, so the results are not actually indicative of much. This additionally explains why even after epoch 20, we see that the model is easily able to separate 177 classes. This occurs in the initial FFNs, and then all future FFNs become uninformative, since this separability is preserved by the high dimensionality of all those FFNs too. To make this even more unreliable, these are tested on the training data of Tiny ImageNet, on which the ViT was shown to get 99.8% accuracy.

Questions for authors

The main thing that could affect my assessment is if my understanding of the theory and required assumptions for this analysis is wrong, and the authors are able to show why the assumptions are much less restrictive than I think they are. Additionally, on the experimental side, either increasing the data or reducing the dimensionality so that the problem isn't trivially separable, and showing strong results in that case, could help.

Reviewer_HhJwOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

The submitted paper presents a Direcitonal Linear Separability measure (DLSM) for machine learning models, with particular attention to neural networks. They present a number of theoretical results showing how this measure is invariant with respect to injective affine transformations. Moreover, they present additional results showing how to improve or certificate DLSM.

Critical strengths
  • The paper addresses a topic central to classical machine learning, namely, linear separability.
  • In particular, the authors address the problem with a sound and consistent theoretical approach.
Critical weaknesses
  1. Clarity: The language employed is convoluted and unnecessarily complicated. In particular, authors introduce notation and language that are not properly specified or described; see Question 2 for specific instances where additional clarification would be benefictial. Furthermore, neither intuitions nor motivations behind these constructions are discussed in the paper.
  2. Identified gap: the authors do not properly highlight the research question and gap that is addressed by the presented manuscript. In the Related Work section they highlight and enumerate differences with analogous previous work but, in my opinion, fail to explain why that gap is relevant to the community.
  3. Reproducibility: Authors do not provide supplementary material, therefore the code for reproducing experiments is unavailable.

The score is mainly influenced by the point 1 and 2 outlined above. Additionally, due to the absence of supplementary materials providing the code of the numerical validations shown in the paper, I am prompted to rise reproducibility concerns.

However, I am open to discussion and to reassess my score.

Questions for authors
  1. Although presenting non-linear decision boundaries, is it possible to extend the analysis to transformers and attention layer?
  2. What is an affine-tube? What is the intuition behind this concept in this context? Is it connected to the notion of margin?
  3. Since their decision boundary is highly nonlinear, why should this method be usefull to describe the behaviour of neural networks or transformers?

Comments on the paper

1 comment

Dear reviewers,

This is a friendly reminder to read the authors’ rebuttals and post your responses by Nov 18. Please briefly indicate which concerns have been addressed and which remain unresolved. Even if your overall assessment remains unchanged, please provide a brief response to the authors. Please also update your reviews and scores as appropriate. Thank you for your time and effort!

AC.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.