acceptodds
Under review as a conference paper at ICLR 2027

Implicit Bias of Gradient Descent under Feature-Mediated Spurious Correlations

Abstract

We study how feature-mediated spurious correlations affect the learning dynamics of gradient descent on linearly separable data. Unlike the widely studied label-mediated setting, where the spurious feature depends only on the class label and is conditionally independent of the causal features, feature-mediated correlations entangle spurious and causal signals at the instance level through group-specific operators, making them both harder to detect in practice and more complex to analyse theoretically. We first generalise the implicit-bias result of soudry2018implicit to continuous distributions. Building on this, we derive closed-form group-wise classification error rates. In the general setting, each group's error decays at a rate governed by its hard-margin . In the isotropic regime, we identify a single exponent that captures the competition between the minority's geometric margin advantage and the coupling induced by the spurious alignment operators. A phase transition occurs at : below it, both groups decay at rate and group proportion directly controls learning speed; above it, the minority escapes the -dependence entirely and converges at the faster polynomial rate . These results formalise in closed form the widely observed phenomenon that group imbalance implicitly biases optimisation dynamics toward the majority, while revealing that the effect of imbalance can be entirely superseded by geometric margin advantages. Finally we establish a non-vanishing error at test time, under spurious correlation shift. Synthetic experiments confirm the rates, the prefactors and the phase transition, and last-layer retraining on Waterbirds shows the predicted scaling of each group’s error on five frozen representations.

Primary area:
learning theory
Keywords:
Gradient Descent, model embedding, spurious correlation

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_rWp3OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This submission investigates the optimization dynamics and implicit bias of gradient descent under feature-mediated spurious correlations on linearly separable data. In contrast to the conventional label-mediated setting.

Critical strengths
  • Conceptual Framing of Feature-Mediated Correlations and Representation Linearisation: The paper tackles an important and under-theorized setting by distinguishing instance-level feature-mediated spurious correlations (s∣r∼Agr+ξs \mid r \sim A_g r + \xi) from classic label-mediated dependencies (s⊥r∣y,gs \perp r \mid y, g). The Sparse Autoencoder (SAE) empirical analysis on Colored-MNIST provides an intuitive perspective on representation learning, demonstrating that multi-layer neural networks can approximately linearize nonlinear input-level correlations within latent feature spaces.

  • Continuous-Distribution Population Implicit-Bias Formulation: Extending the separable implicit-bias framework of Soudry et al. (2018) from discrete support vectors to continuous distributions with strictly positive boundary margin densities is a substantive theoretical formulation. Formulating population gradient descent via Laplace-type integrals over continuous margin levels and controlling the transverse parameter drift via an orthogonal energy argument (Lemma I.8) offers an insightful mathematical explanation for the emergence of the −w^ln⁡ln⁡zt-\hat{w} \ln\ln z_t residual correction.

Critical weaknesses
  • Mathematical Breakdown at the Critical Threshold (α=1\alpha = 1) and Invalid Asymptotic Derivations: The paper's central theoretical claim of a continuous phase transition at α=1\alpha = 1 is mathematically flawed:
  1. Failure at α=1\alpha = 1: Theorem D.4(ii) (Eq. 43) groups α≥1\alpha \ge 1 together and claims a minority rate of Θ((ln⁡zt)α−1/ztα)\Theta((\ln z_t)^{\alpha-1} / z_t^\alpha), which evaluates to Θ(1/zt)\Theta(1/z_t) at α=1\alpha = 1. However, the proof in Appendix E.5 explicitly relies on Eq. (90): Δ˙≤(1−ϵ)(1+μA)(1−α)Imaj≤−C/z\dot{\Delta}\le (1-\epsilon)(1+\mu_A)(1-\alpha)\mathcal{I}_\text{maj}\le -C/z. At α=1\alpha = 1, the leading coefficient (1−α)(1-\alpha) vanishes identically, which invalidates the estimate and breaks the subsequent integration. Analyzing the underlying ODE system at α=1\alpha = 1 reveals that the trajectory gap satisfies u:=γ~minb−a∼ln⁡ln⁡zu := \tilde{\gamma}_{\text{min}} b - a \sim \ln\ln z, which induces an additional logarithmic factor, yielding a true critical rate of Θ(1/(ztln⁡zt))\Theta(1 / (z_t \ln z_t)) rather than Θ(1/zt)\Theta(1/z_t).

  2. Invalid Application of Karamata's Theorem: Proposition H.5 applies Karamata’s theorem to zδ+ϵ(z)z^{\delta + \epsilon(z)} under the assumption ϵ(z)→0\epsilon(z) \to 0. This inference is mathematically invalid: ϵ(z)→0\epsilon(z) \to 0 does not guarantee slow variation (e.g., ϵ(z)=sin⁡(ln⁡z)/ln⁡z→0\epsilon(z) = \sin(\ln z)/\sqrt{\ln z} \to 0, yet zϵ(z)=exp⁡(ln⁡zsin⁡(ln⁡z))z^{\epsilon(z)} = \exp(\sqrt{\ln z}\sin(\ln z)) oscillates unbounded), undermining the noisy asymptotic rates in Appendix F.2.

  3. Flawed Asymptotic Exponent Notation: In Theorem 5.1 and Corollary G.1, the error bounds are expressed as Θ((ln⁡zt)kztγg+o(1))\Theta\left(\frac{(\ln z_t)^k}{z_t^{\gamma_g + o(1)}}\right). Stating an exact logarithmic power in the numerator while leaving an unrefined zto(1)z_t^{o(1)} factor in the denominator is inconsistent, as zto(1)z_t^{o(1)} can absorb or dwarf any polylogarithmic factor (ln⁡zt)k(\ln z_t)^k.

  • Artificial 1D Dimensional Collapse, Generative Model Conflict, and Metric Misinterpretation:
  1. Subspace Collapse: To obtain closed-form ODE dynamics, the analysis imposes simultaneous isotropy (Definition 5.2), zero orthogonal initialization (Πˉvψ0=Πˉvϕ0=0\bar{\Pi}_v \psi_0 = \bar{\Pi}_v \phi_0 = 0, Assumption D.2), and conditional zero-mean orthogonal features (E[r⊥∣Z,g]=0\mathbb{E}[r_\perp \mid Z, g] = 0, Assumption D.1). This artificially restricts parameter updates to a 1D scalar line Spanv\text{Span}\\{ v \\}, bypassing the multi-dimensional transverse optimization dynamics of neural network training. Under non-zero feature noise ξ\xi, updates inject non-zero orthogonal components via Ag⊤E[qtξ]A_g^\top \mathbb{E}[q^t \xi] that are not bounded or tracked in Appendix F.2.

  2. Generative Model Inconsistency: Definition 2.3 (Eq. 3) defines the mixture conditioned on rr with constant probabilities 1−ϵ1-\epsilon and ϵ\epsilon (P(G=min∣r)=ϵP(G=\text{min} \mid r) = \epsilon). This implies G⊥rG \perp r, forcing identical rr-margin distributions across both groups (Law(r∣Gmaj)=Law(r∣Gmin)\text{Law}(r \mid \mathcal{G}_{\text{maj}}) = \text{Law}(r \mid \mathcal{G}_{\text{min}})), which directly contradicts the subsequent assumption of unequal group margins (γ~min>γ~maj=1\tilde{\gamma}_{\text{min}} > \tilde{\gamma}_{\text{maj}} = 1).

  3. Conflation of Surrogate Loss with 0–1 Error: Throughout the manuscript, the soft surrogate error Eg[1−pyt]=Eg[σ(−yw⊤x)]\mathbb{E}_g[1 - p_y^t] = \mathbb{E}_g[\sigma(-y w^\top x)] is referred to as "classification error". On linearly separable distributions with positive margin, the actual 0–1 misclassification error Pr⁡g(yw⊤x≤0)\Pr_g(y w^\top x \le 0) becomes identically zero in finite time. Asserting minority learning bottlenecks or worst-group performance guarantees based purely on decaying soft loss tails overstates the practical robustness implications.

Questions for authors
  1. Resolution of the Critical Phase Transition at α=1\alpha = 1: In Eq. (90), the bound Δ˙≤(1−ϵ)(1+μA)(1−α)Imaj≤−C/z\dot{\Delta} \le (1-\epsilon)(1+\mu_A)(1-\alpha)\mathcal{I}_{\text{maj}} \le -C/z requires 1−α>01-\alpha > 0, failing when α=1\alpha = 1. Analyzing the coupled drift equations at α=1\alpha = 1 shows that u=γ~minb−a∼ln⁡ln⁡zu = \tilde{\gamma}_{\text{min}} b - a \sim \ln\ln z, which implies an asymptotic minority rate of Θ(1/(ztln⁡zt))\Theta(1 / (z_t \ln z_t)). How do you reconcile this additional logarithmic factor with the claimed rate of Θ(1/zt)\Theta(1/z_t) and the continuity claim at α=1\alpha = 1 in Theorem D.4 and Theorem 5.3?

  2. Generative Model Specification vs. Margin Discrepancy: In Definition 2.3 (Eq. 3), the group assignment is specified conditionally on rr with fixed probability ϵ\epsilon, which forces rr and GG to be independent and implies identical conditional rr-distributions. How can this formulation support distinct margin supports γ~min>γ~maj=1\tilde{\gamma}_{\text{min}} > \tilde{\gamma}_{\text{maj}} = 1 and group-specific densities λmaj≠λmin\lambda_{\text{maj}} \ne \lambda_{\text{min}} without reformulating the joint law as G∼Bernoulli(ϵ)G \sim \text{Bernoulli}(\epsilon) and r∣G∼DGr \mid G \sim \mathcal{D}_G?

  3. Rigorous Foundation for Proposition H.5: Proposition H.5 invokes Karamata's theorem based solely on the assumption that ϵ(z)→0\epsilon(z) \to 0. Since ϵ(z)→0\epsilon(z) \to 0 does not guarantee that zϵ(z)z^{\epsilon(z)} is slowly varying, what additional regularity conditions (such as bounded variation or explicit decay rates on ϵ′(z)\epsilon'(z)) are required to ensure that Eq. (187) and the resulting Θ(1/z)\Theta(1/z) rates hold?

  4. Transverse Drift under Feature Noise: In Eq. (122), the parameter updates include the term Ag⊤E[qtξ]A_g^\top \mathbb{E}[q^t \xi]. Because the noise ξ\xi is arbitrary and bounded within a ball in Rds\mathbb{R}^{d_s}, these gradient updates generally possess non-zero projections onto (v)⊥(v)^\perp. Why does this transverse drift not invalidate the reduction to a 1D scalar ODE in Lemma F.5?

  5. Practical Interpretation of Error Metric: Given that the 0–1 classification error Pr⁡g(yw⊤x≤0)\Pr_g(y w^\top x \le 0) reaches zero on linearly separable data, what exact conclusions can be drawn from the asymptotic decay of the surrogate loss Eg[1−pyt]\mathbb{E}_g[1 - p_y^t] regarding actual test-time worst-group classification accuracy?

Reviewer_w5TuOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how spurious correlations affect the optimization dynamics of GD. They theoretically analyzed a specific problem in a linearly separable population classification setting. Suppose there are relevant features (r) and derived (spurious) features (s) that are generated from r. Suppose also the mapping from r to s is different for a majority and a minority subpopulation. When gradient descent minimizes population logistic risk, how quickly does its error improve for the majority group versus the minority group?

The authors find that learning speed is determined by 3 factors:

  1. group size;
  2. the group’s margin to the decision boundary;
  3. the group-specific mappings used to derive feature (s).


They describe this as a phase transition controlled by a parameter \alpha that controls how factors 2 and 3 are weighted: Below \alpha = 1 the group proportion sets the learning speed of each group; above it, the margin geometry. Synthetic experiments are used to verify the theory.

Critical strengths

I think conceptually, how GD works when there is spurious correlation is an interesting question that is worth analyzing theoretically on specific small examples to see if it aligns with the trend we can observe in practice, and what specific factors determine which groups are learned faster.

Critical weaknesses

I find the weakness is mainly in two aspects:

  1. The clarity needs significant improvement. It’s currently significantly limiting the message being conveyed.
  2. The paper uses a rather specific setting, and it’s not properly explained why this framework makes sense or is practically relevant. I did not examine the correctness of the proofs in the appendix.

More details:

Novelty

  • The core assumption is feature-mediated spurious correlation, as opposed to label-mediated spurious correlation. Is this an existing concept, or something proposed by the authors that is potentially a contribution itself? This was unclear. Because overall, I find that saying this type of feature is “spurious” is already a strong statement. For example, in the right panel of Figure 1, it is unclear what the task label (y) is. Why, then, is feature (s) considered spurious, or, more generally, how is a derived feature determined as spurious?

Significance

  • While it is reasonable to do theoretical analysis on a constructed narrow example, I find there to be limited practical motivation or evidence that this setup occurs in practice and why it is important. Section 4 briefly discusses possible practical settings, including last-layer retraining and sensor functions, but these are mentioned rather than carefully explained.
  • It is unclear whether the results are predictive enough in finite-sample neural-network training to make the phenomenon broadly important. Although this is mentioned in the limitations section.
  • The experiments are limited in scale and scope, mostly sanity checks of the theory, with waterbird one being the only real data experiment, and that also verifies only the simplest setting. It’s unclear how the theory related to alpha acts out in this experiment. It validates that GD learns a group of less total loss weight slower, but does not seem to validate the distinctive feature-mediated spurious feature mechanism. The test-shift failure theorem also doesn’t seem to map to an experiment.

Clarity

  • The abstract introduces many concepts without defining them adequately. At first glance, it’s very difficult to understand.
  • Many mathematical symbols mentioned in the abstract, parts of the writing of theorems and proof ideas, are either undefined, defined only later, or defined only in the appendix.
  • The theory is poorly presented and difficult to follow. The main message is not clear from the theorems alone. Overall, I find the mathematical results to be overly complicated with unnecessary notations. They might be necessary for the full proof in the appendix, but for the main paper, the goal should be to explain the main idea and findings with notations and language as simple and clear as possible. The theorems should foreground the important dependence on group proportion, margin, and spurious alignment. At present, the expressions inside the (\Theta) notation make the reader search for the important difference between cases rather than explaining the result directly. Given this, it’s also quite possible I’ve missed or misunderstood some findings.
Questions for authors

(same as some points in weakness)

  • Is feature-mediated spurious correlation an existing concept?
  • Why is the feature (s) considered spurious in Figure 1?
  • Could the main result that GD dynamics are affected by both group size and margin occur in the label-mediated regime? Or even without spurious features at all, when there is some other notion of groups? Are those existing results?
Reviewer_FWg7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies population gradient descent (GD) on linearly separable data with a fixed representation. It separates the representation into causal features (rr) and spurious features (ss), with s=Ar+ξs = Ar + \xi in the majority group, and s=Br+ξs = Br + \xi in the minority group. The minority has proportion ϵ\epsilon. Therefore, the spurious feature vector depends on the causal feature vectors through group-specific linear transformations. Under separability, this paper shows that GD converges to the population minimum-norm separator (w^\hat{w}) for continuous and compactly supported distributions. This generalizes the finite-sample result of Soudry et al. Moreover, the paper further shows that the prediction error is group specific and is governed by the group margin under w^\hat{w}. Finally, the paper provide experimental results to support its theoretical claims.

Critical strengths

The extension of the Soudry’s finite-sample result to the population level is meaningful. The group-level analysis and results can provide more insights on the dynamics of GD.

Soudry et al., The Implicit Bias of Gradient Descent on Separable Data

Critical weaknesses

Based on the model specified in (3), it suggests that the group membership is independent of the causal feature vector rr. Consequently, the groups have the same causal-feature distribution. This implies that γmaj=γmin\gamma_{maj} = \gamma_{min}. By eqn (21) (min⁡(γmaj,γmin)=1\min(\gamma_{maj}, \gamma_{min}) = 1), it holds that γmaj=γmin=1\gamma_{maj} = \gamma_{min} = 1. Based on eqn (10) and the definition of α\alpha, it then holds that α<1\alpha < 1. This means that the model rules out the α>1\alpha > 1 regime. This seems to be a critical technical flaw. Can the authors address this?

The paper is confusing and difficult to follow. Various quantities and assumptions in its theorem statements are defined only in the appendix. It is important to clearly define all the constants and state the assumptions used in the analysis.

Questions for authors

See Weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.