🤖 machine learning

Mitigating Spurious Correlations with Memorization-Guided Dataset De-Biasing

This paper proposes a two-stage sample scoring function that disentangles the learning dynamics of core and spurious features to identify and prioritize informative data subsets, enabling standard models to achieve superior debiasing performance with as little as 10% of the original training data without requiring group labels.

Original authors: Arda Fazla, Abolfazl Hashemi

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Arda Fazla, Abolfazl Hashemi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "Cheat Code" of AI

Imagine you are teaching a child to identify birds. You show them 100 pictures.

90 pictures show woodpeckers standing on trees.
10 pictures show woodpeckers standing on water (maybe they are fishing).

If you ask the child, "What is this bird?" while pointing at a woodpecker on a tree, they will likely say "Woodpecker." But here is the trap: The child isn't actually looking at the bird's beak or feathers (the core features). They are looking at the tree (the spurious feature).

Because the tree appears in 90% of the woodpecker photos, the child learns a "cheat code": If I see a tree, it's a woodpecker.

Now, show the child a picture of a woodpecker on a lake. Because the child is relying on the "tree cheat code," they get confused and might say, "That's not a woodpecker!" or guess the wrong bird entirely.

In the world of AI, this is called a spurious correlation. The AI learns the easy, obvious pattern (the background) instead of the hard, important pattern (the actual object). This leads to the AI failing when it encounters something that breaks the pattern (like a bird on water).

The Old Solution vs. The New Problem

Researchers have tried to fix this by throwing away the "easy" photos and keeping only the "hard" ones (the birds on water) to force the AI to learn the real features.

However, there is a catch: To do this, you usually need to know which photos are "hard" and which are "easy." This requires a human to label every single photo with a secret tag (e.g., "This is a bird on water"). In the real world, we rarely have these labels.

Furthermore, the paper argues that the tools we usually use to find "hard" photos are broken. They look at the AI's mistakes and say, "This photo is hard because the AI got it wrong." But if the AI is already cheating using the background, it will get the "easy" photos (birds on trees) wrong only if the background is missing, and get the "hard" photos right by accident. The old tools get confused and pick the wrong photos to keep.

The New Solution: The "Two-Stage Detective"

The authors propose a new method called TCSL-CS. Think of it as a two-step detective process that doesn't need human labels to find the "cheaters."

Step 1: The "Bias Detective" (Training the Spurious Model)

First, they train a special AI model designed to be bad at looking at the bird but great at looking at the background.

They train this model, but they give it a special reward: "If you get the answer right just by looking at the background, you get a bonus point."
This forces the model to learn the "cheat code" (the background) as fast as possible.
Once this model is trained, they look at the photos. If the "Bias Detective" is very confident about a photo, it means the photo has a strong, obvious background pattern. If the detective is confused, it means the background is tricky or missing.

Step 2: The "Truth Detective" (Training the Core Model)

Next, they train a second model to learn the real bird features.

Here is the clever part: They tell this second model, "Ignore the background. In fact, we already know what the background looks like (from Step 1), so we will subtract that knowledge from your task."
This forces the second model to focus entirely on the bird's beak and feathers because the background "noise" has been canceled out.

Step 3: The Scorecard (TCSL)

Now, the authors have two scores for every single photo:

Spurious Score: How much did the photo rely on the background? (From Step 1).
Core Score: How hard was it to identify the bird once the background was ignored? (From Step 2).

They use these scores to build a new, smaller training set (a Coreset). They pick photos that:

Have low Spurious Scores (meaning the background isn't a cheat code).
Have high Core Scores (meaning the bird is actually hard to identify, so the AI needs to learn it well).

The Result: A Smaller, Smarter Dataset

The authors tested this on several datasets (like identifying birds on land vs. water, or cats vs. dogs in different environments).

The Magic: They were able to throw away 90% of the original training data.
The Outcome: They took the remaining 10% of the "best" photos (selected by their Two-Stage Detective) and trained a standard AI on them.
The Win: This tiny, curated dataset made the AI perform better than complex, expensive AI methods that require human labels. The AI stopped cheating with backgrounds and started actually recognizing the objects, even in difficult situations.

Summary Analogy

Imagine you are preparing for a history exam.

The Old Way: You study every single page of the textbook. You memorize that "Page 50 has a picture of a red flag," so you guess "Red Flag" for every question. You fail when the question is about a blue flag.
The Paper's Way: You hire a tutor who first teaches you to spot the "Red Flag" trick (Step 1). Then, the tutor tells you, "Now, ignore the flags and study the actual dates and names" (Step 2).
The Result: The tutor gives you a cheat sheet with only the 10 most important, tricky questions (Step 3). You study just those 10 questions, and you pass the exam with flying colors, knowing the real history, not just the pictures.

The paper claims that by using this "Two-Stage" approach to select the best 10% of data, you can build a robust AI without needing to know the secret labels or use complex math tricks.

Technical Summary: Mitigating Spurious Correlations with Memorization-Guided Dataset De-Biasing

Problem Statement

Real-world datasets frequently contain spurious correlations—features that are highly consistent within a class but not causally related to the target label. Deep learning models trained via Empirical Risk Minimization (ERM) often exhibit simplicity bias, learning these simpler spurious features before capturing complex core features. This leads to poor worst-group accuracy (WGA) on test data where spurious correlations do not hold, particularly for minority groups where the spurious attribute disagrees with the label.

While existing approaches to mitigate this bias often deviate from standard ERM by employing complex optimization techniques or requiring access to group labels (spurious attributes), there is a need for methods that can reduce reliance on spurious correlations while retaining standard ERM training. A potential solution lies in coreset selection—constructing high-quality subsets of data that represent the full training distribution. However, the paper argues that widely used sample scoring functions (e.g., EL2N, SelfSup) and coreset selection algorithms fail in this context. Due to simplicity bias, these scoring functions are dominated by spurious features, misclassifying samples with strong spurious correlations as "easy" and samples without them as "hard." Consequently, standard selection strategies fail to prioritize the difficult core features necessary for group robustness.

Methodology

The authors propose TCSL-CS (Two-Stage Cumulative Sample Loss-guided Coreset Selection), a framework designed to disentangle the learning dynamics of core and spurious features without requiring group labels.

1. Theoretical Foundation

The paper provides a theoretical analysis under the Neural Tangent Kernel (NTK) regime and a spiked covariance model. It demonstrates that under the simplicity bias condition ( $\beta_s(2\alpha - 1) > \beta_c$ , where $\beta$ represents feature strength and $\alpha$ the majority group fraction), spurious features are learned significantly faster than core features during the early stages of training. This rapid learning causes sample losses (and thus standard scores) to be dominated by the spurious attribute, rendering them ineffective for identifying samples with difficult core features.

2. Two-Stage Model Training

To decouple these dynamics, the authors introduce a two-stage training scheme (Algorithm 1):

Stage 1 (Spurious Model, $f_s$ ): A biased model is trained to learn only spurious features. This is achieved by upweighting samples with the lowest loss at each epoch, effectively amplifying the learning of samples where the spurious attribute aligns with the label ( $a=y$ ).
Stage 2 (Core Model, $f_c$ ): The weights of $f_s$ are frozen. A core model is trained using a modified loss function that incorporates the logit of the spurious model as a negative feedback mechanism:
$L(B_t, W^c_t) = \sum_{i \in B_t} \lambda_i \ell\left(y_i, f_c(x_i; W^c_t) + \log(f_s(x_i))\right)$
This pushes $f_c$ to learn the core features by compensating for the predictions of the biased model.

3. Two-Stage Cumulative Sample Loss (TCSL)

Based on the trained models, the authors define a new scoring metric, TCSL, which consists of two components:

$TCSL_s(x)$ : The cumulative loss of the spurious model $f_s$ . High scores indicate samples where the spurious feature is weak or misaligned with the label.
$TCSL_c(x)$ : The cumulative loss of the core model $f_c$ . This score reflects the difficulty of the core features, independent of the spurious attribute.

4. TCSL-CS Algorithm

The coreset selection algorithm (Algorithm 2) operates in two stages to construct a robust subset:

Identify High-Spuriosity Samples: Using wKMeans on $TCSL_s$ scores and loss curves, the algorithm identifies a group of samples ( $G_{high}$ ) with high spurious difficulty (i.e., samples lacking strong spurious correlations). A subset of these is selected to ensure the coreset includes challenging minority-like samples.
Fill Remaining Quota: The remaining slots are filled from the majority group ( $G_{low}$ $G_{l o w}$ ) based on $TCSL_c$ $T C S L_{c}$ scores.
- If the selection ratio $r$ is high ( $r \geq \tau$ ), the algorithm selects samples with the lowest $TCSL_c$ scores (easiest core features) to maximize average accuracy.
- If $r$ is low ( $r < \tau$ ), the algorithm uses a histogram-based selection (SelectHist) to sample from bins of varying $TCSL_c$ difficulty, ensuring diversity in core feature difficulty and preventing the exclusion of hard core features.

Key Contributions

TCSL Score: A novel metric that separately quantifies the learning difficulty of core and spurious features, addressing the failure of existing scores to disentangle these components under simplicity bias.
TCSL-CS Algorithm: A group-robust coreset selection method that achieves high worst-group accuracy without requiring access to group labels. It leverages the TCSL score to construct coresets that balance average accuracy and worst-group performance.
Theoretical Analysis: A rigorous analysis under the NTK regime proving that spurious features dominate early learning dynamics, explaining why standard loss-based scores fail in spurious settings.
Empirical Superiority: Demonstrating that a standard ERM model trained on the TCSL-CS coreset outperforms state-of-the-art debiasing techniques (including those requiring group labels) while using only 10% of the original training data.

Experimental Results

The authors evaluated TCSL-CS on four datasets known for spurious correlations: Waterbirds, cMNIST, MetaShift, and UrbanCars-B.

Performance vs. Debiasing Baselines: On Waterbirds, TCSL-CS improved the WGA of a standard ERM model by 10.44%, outperforming GroupDRO (which requires group labels) and other debiasing methods like LC, DFR, and JTT. On cMNIST (where spurious correlations are extremely strong, $\alpha=0.995$ ), TCSL-CS surpassed the next best method by over 11%.
Performance vs. Coreset Selection Baselines: TCSL-CS significantly outperformed standard coreset selection methods (EL2N, SelfSup, Random, D2) across all datasets. For instance, on Waterbirds, EL2N-based methods achieved WGA scores as low as 39.88%, whereas TCSL-CS achieved 91.91%.
Data Efficiency: The method achieved state-of-the-art performance using only 10% of the training data.
Ablation Studies: Experiments confirmed that $TCSL_s$ and $TCSL_c$ successfully disentangle spurious and core feature difficulties, showing high cosine similarity with scores derived from models trained exclusively on spurious or core features, respectively.

Significance and Claims

The paper claims that TCSL-CS offers a scalable and principled approach to debiasing. Unlike many existing methods that require complex optimization or group labels, TCSL-CS:

Retains Standard ERM: It allows the use of standard ERM training on the selected coreset, avoiding the need for specialized loss functions or optimization procedures during the final training phase.
Reduces Computational Cost: By selecting a small subset (e.g., 10%) of data, it significantly reduces the computational burden of training large models, as the expensive coreset selection step is performed once, and the resulting subset can be reused.
Generalizability: The selected coresets are architecture-agnostic, yielding consistent performance gains across different model backbones (ResNet, Inception, DenseNet).

The authors position their work as the first coreset selection algorithm specifically designed to achieve both high average and high worst-group accuracy in the absence of group labels, effectively mitigating the negative impact of simplicity bias on dataset selection.