💻 computer science

POUR: A Provably Optimal Method for Unlearning Representations via Neural Collapse

This paper introduces POUR, a provably optimal geometric projection method for machine unlearning that removes specific visual concepts at the representation level by leveraging Neural Collapse theory to balance forgetting efficacy, retention fidelity, and class separation, thereby outperforming existing state-of-the-art approaches.

Original authors: Anjie Le, Can Peng, Yuyuan Liu, J. Alison Noble

Published 2026-04-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Anjie Le, Can Peng, Yuyuan Liu, J. Alison Noble

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

Imagine you have a highly trained librarian who has memorized the entire contents of a massive library. One day, a legal requirement (like the "Right to be Forgotten") demands that the librarian completely forget everything about a specific book, say, a cookbook, without retraining their entire brain from scratch.

Most current methods try to solve this by simply telling the librarian, "When someone asks about the cookbook, just guess randomly or say you don't know." But the problem is that the librarian's brain still contains the detailed recipes and images of the cookbook deep inside. If you ask the right questions, they might accidentally reveal those forgotten details. This is like trying to hide a book by putting a "Do Not Read" sign on it, while the book remains open on the shelf.

The paper "POUR" proposes a smarter, more surgical solution. It doesn't just change the librarian's answers; it actually reorganizes the librarian's mental map of the library to physically remove the cookbook's section while keeping the rest of the library perfectly organized.

Here is how they did it, broken down into simple concepts:

1. The "Perfect Arrangement" (Neural Collapse)

The authors noticed that when AI models are trained well, the way they organize information isn't messy. It's like a perfectly symmetrical geometric shape. Imagine all the different types of books (cats, dogs, cars, etc.) are represented by points in space. In a well-trained model, these points arrange themselves like the corners of a perfect, symmetrical pyramid (mathematically called a Simplex Equiangular Tight Frame). Every category is equidistant from every other category, creating a perfectly balanced structure.

2. The "Surgical Cut" (The POUR Method)

The core idea of POUR is to use this perfect geometry to perform a "surgical cut."

The Problem: If you just delete the "cookbook" corner of the pyramid, the remaining corners might collapse or get messy, ruining the librarian's ability to distinguish between the other books.
The Solution: The authors discovered a mathematical trick. If you have this perfect pyramid and you "project" (or shine a light on) the remaining corners onto a flat surface that ignores the cookbook corner, the remaining corners automatically form a new, smaller, but still perfectly symmetrical pyramid.

Think of it like a 3D sculpture of a star. If you carefully slice off one point of the star and look at the remaining shape from a specific angle, the remaining points still form a perfect, balanced shape. The authors call this POUR (Provably Optimal Unlearning of Representations). It's "provably optimal" because math proves that this is the only way to remove the specific information while keeping the rest of the structure perfectly intact.

3. Two Ways to Do It

The paper offers two versions of this "surgical cut":

POUR-P (The Instant Cut): This is a one-time mathematical operation. You take the model's existing knowledge and instantly apply the projection. It's like taking a photo of the library and digitally erasing the cookbook section in a single click. It's fast and doesn't require retraining.
POUR-D (The Guided Cleanup): This is a bit more thorough. It uses the "Instant Cut" model as a teacher and trains the original model to mimic this new, cleaner version. It's like having the librarian practice looking at the library through the new "lens" to ensure they truly forget the cookbook and don't accidentally remember it later.

4. How Do We Know It Worked? (The Scorecard)

The authors created a new score called RUS (Representation Unlearning Score).

Old methods were like checking if the librarian says "I don't know" when asked about the cookbook. But the librarian might still be thinking about the recipes.
POUR's method checks the librarian's brain structure. It verifies that the mental "map" of the cookbook has been completely erased, while the maps for cats, dogs, and cars remain sharp and distinct.

The Results

The team tested this on various datasets (like images of animals and medical scans). They found that:

Total Erasure: The model completely forgot the target category. When asked about it, the model didn't just guess; the internal "signal" for that category vanished entirely.
Perfect Retention: The model didn't get confused about the other categories. It remembered them just as well as before, or even better, because the "noise" of the forgotten category was removed.
Better than the Rest: Compared to other methods that just tweak the final answers, POUR actually cleaned up the internal memory, making it much harder for anyone to "reverse engineer" the forgotten information.

Summary

In short, POUR is a method that treats machine unlearning not as a game of "guessing wrong," but as a geometric surgery. By understanding that AI models organize data in perfect, symmetrical shapes, the authors found a way to surgically remove a specific piece of knowledge while ensuring the rest of the structure remains perfectly balanced and functional. It's the difference between telling a student to "pretend they don't know the answer" versus actually removing the chapter from their textbook so they can't possibly remember it.

1. Problem Definition

Machine Unlearning aims to remove the influence of specific training data (e.g., a specific class or subset) from a trained model without retraining from scratch. This is critical for regulatory compliance (e.g., GDPR's "right to be forgotten") and safety in sensitive domains.

The Core Challenge:
Existing methods primarily focus on weak unlearning, which ensures that the final prediction probabilities (logits) of the unlearned model are indistinguishable from a retrained model. However, recent studies indicate that these methods often leave internal feature representations intact. Consequently, forgotten information can still be recovered via linear probing or feature inversion, leading to privacy leakage. There is a lack of methods that guarantee forgetting at the representation level (within the feature extractor) while preserving the geometric structure required for the remaining classes.

2. Methodology: POUR

The authors propose POUR (Provably Optimal Unlearning of Representations), a geometric projection method grounded in the theory of Neural Collapse (NC).

A. Theoretical Foundation: Neural Collapse & ETF

The method relies on the observation that deep classifiers trained with cross-entropy loss converge to a Neural Collapse state, where:

Class features collapse to their class means.
Class means form a Simplex Equiangular Tight Frame (ETF): a set of vectors that are maximally separated and symmetrically arranged in the feature space.
Classifier weights align with these class means.

The authors establish two critical theoretical properties:

Bayes Optimality: The Simplex ETF geometry is not just a training artifact but a sufficient condition for Bayes optimality in balanced classification under isotropic Gaussian assumptions.
Projection Invariance: The orthogonal projection of a Simplex ETF (removing one vertex) onto the orthogonal complement of that vertex yields a smaller Simplex ETF. This means removing a class via projection preserves the optimal geometric separation of the remaining classes.

B. The POUR Algorithm

POUR offers two variants to achieve representation-level unlearning:

POUR-P (Closed-Form Projection):
- Mechanism: Applies an orthogonal projection operator $P = I - \frac{w_u w_u^\top}{\|w_u\|^2}$ directly to the feature space (penultimate layer), where $w_u$ is the weight vector of the class to be forgotten.
- Effect: This projects the features of the forgotten class to the origin (zeroing them out) while mapping the features of retained classes into a lower-dimensional subspace that maintains the Simplex ETF structure.
- Advantage: Instantaneous, no retraining required, and mathematically proven to be optimal.
POUR-D (Projection-Guided Distillation):
- Mechanism: Uses the projected model (POUR-P) as a "teacher." A student model (feature extractor) is fine-tuned on the forget set only to minimize the $L_2$ distance between its features and the projected teacher features.
- Effect: Propagates the forgetting deeper into the feature extractor, ensuring the internal representations align with the optimal ETF geometry of the retained classes.

C. New Metric: Representation Unlearning Score (RUS)

To evaluate unlearning beyond logits, the authors introduce RUS, which quantifies the trade-off between:

Forgetting Efficacy: How close the unlearned model's features are to the retrained model on the forget set.
Retention Fidelity: How well the unlearned model preserves features for the retain set.
Class Separation: How well the retained classes remain separated.
RUS uses Centered Kernel Alignment (CKA) to measure representation similarity, making it robust to random initialization and rotation.

3. Key Contributions

Representation-Level Reformulation: Shifts the unlearning paradigm from output logits to internal feature distributions, defining "Representation-Level Weak Unlearning."
Theoretical Guarantees: Proves that orthogonal projection of a Simplex ETF preserves the ETF structure, providing a provably optimal forgetting operator that maintains Bayes optimality for retained classes.
RUS Metric: Introduces a principled metric to quantify representation-level forgetting and retention, addressing the limitations of existing evaluation protocols.
POUR Algorithm: Proposes a closed-form (POUR-P) and a distillation-based (POUR-D) method that outperforms state-of-the-art baselines.

4. Experimental Results

The authors evaluated POUR on CIFAR-10, CIFAR-100, and PathMNIST (medical imaging), comparing it against baselines like Gradient Ascent, Boundary Shrink, DELETE, and Fine-tuning.

Classification Metrics: POUR achieves near-perfect forgetting (accuracy on forget set $\approx 0$ ) while maintaining high accuracy on retained classes, often outperforming methods that require access to the retain set.
Representation Metrics (RUS & CKA): POUR significantly outperforms all baselines in RUS scores. While other methods (like Gradient Ascent) disrupt the geometry of retained classes, POUR preserves the ETF structure, resulting in feature distributions indistinguishable from a retrained model.
Domain Shift (PathMNIST): In a setting with domain shift (internal vs. external test sets), POUR demonstrated robust generalization, whereas methods like "Random Label" failed to generalize, suggesting they merely "masked" the forget set rather than erasing knowledge.
Cross-Modal & Segmentation: The method was successfully extended to CLIP (text-to-image unlearning) and Semantic Segmentation (VOC2012), demonstrating its generality beyond standard classification.

5. Significance

Privacy & Security: By ensuring forgetting occurs at the representation level, POUR mitigates the risk of information leakage through feature inversion, a critical vulnerability in current unlearning methods.
Theoretical Insight: It bridges the gap between Neural Collapse theory and practical unlearning, showing that the geometric structure of deep networks can be manipulated mathematically to achieve optimal forgetting without retraining.
Efficiency: The closed-form nature of POUR-P allows for instantaneous unlearning, making it highly scalable for large foundation models where retraining is infeasible.

In summary, POUR provides a geometrically principled, theoretically optimal, and empirically superior framework for machine unlearning, ensuring that models truly "forget" specific concepts while retaining their utility for other tasks.