🤖 machine learning

Bilinear autoencoders find interpretable manifolds

This paper introduces bilinear autoencoders that utilize quadratic latents to uncover interpretable, multi-dimensional manifolds in neural network activations, demonstrating that nonlinear geometric priors can systematically improve reconstruction and reveal composite concepts that linear methods miss.

Original authors: Thomas Dooms, Ward Gauderis, Geraint Wiggins, Jose Oramas

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Thomas Dooms, Ward Gauderis, Geraint Wiggins, Jose Oramas

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 are trying to understand how a massive, complex machine (like a modern AI language model) thinks. For a long time, researchers have tried to do this by looking for straight lines. They assumed that if a concept exists inside the AI, it's like a single arrow pointing in a specific direction. If the AI thinks about "cats," there's a specific line in its brain that lights up.

This paper argues that this "straight line" view is too simple. Instead, the paper suggests that many concepts in AI are more like shapes, bubbles, or curved surfaces. To find these shapes, the authors built a new tool called a Bilinear Autoencoder.

Here is a breakdown of their ideas using simple analogies:

1. The Problem: The "Straight Line" Map is Missing Things

Think of the AI's brain as a giant, multi-dimensional room filled with invisible furniture (concepts).

Old Method (Linear Autoencoders): Imagine trying to map this room using only a ruler. You can measure straight lines and flat walls. If a concept is a "straight line," you find it easily. But if a concept is a circle, a sphere, or a curved hill, a ruler can't describe it well. You end up needing hundreds of tiny, broken ruler segments to approximate a circle, which is messy and confusing.
The Reality: The authors found that many AI concepts are curved. For example, the concept of "years" (like 1990, 2000, 2010) isn't just a line; it's a specific geometric shape. The concept of "U.S. locations" isn't a single point; it's a scattered cloud of points that form a specific cluster.

2. The Solution: The "Shape-Shifting" Lens

The authors created a new tool that doesn't just look for straight lines; it looks for curved shapes (mathematically called "quadrics," like ellipsoids or hyperbolas).

The Analogy: Imagine the old tool was a flashlight that only shines a straight beam. The new tool is a laser cutter that can carve out curved shapes.
How it works: Instead of asking, "Is this input on the line?", the new tool asks, "Is this input inside this curved bubble?" or "Is this input on this specific curved surface?"
The "Bilinear" part: This is just a fancy way of saying the tool looks at how two things interact. It doesn't just look at "Cat"; it looks at the relationship between "Cat" and "Meow" simultaneously to define the shape of the concept.

3. What They Found: Curved Shapes are Everywhere

When they used this new tool on a language model (Qwen 3.5), they found three main types of "shapes" that the old tools missed:

The "Slab" (Simple Curve): Imagine a thick sandwich. The concept activates if you are between two parallel slices of bread, regardless of which side of the bread you are on.
- Example: The concept of "years" (19xx, 20xx). The tool found a shape that captures all years, ignoring whether the number is positive or negative in a mathematical sense.
The "Cluster" (Bubbles): Imagine a bunch of grapes. The concept isn't one big shape, but a group of distinct points that form a cluster.
- Example: The concept of "U.S. locations." The tool found a shape that groups "U.S.", "United", and "America" together, while ignoring "London" or "Paris," even though they are all locations.
The "Saddle" (Complex Curve): Imagine a horse saddle. You can go up in one direction and down in another.
- Example: The phrase "for instance." The tool learned to activate when it sees "for instance" or "e.g.," but deactivate (turn off) when it sees the word "for" in other contexts (like "for the love of"). It captures the nuance of the context, not just the word itself.

4. Why This Matters (The "Dictionary" Analogy)

The authors compare their method to building a dictionary of concepts.

Old Dictionary: You have thousands of words, but they are all just straight lines. To describe a circle, you have to use 50 different words that are slightly off-center. It's inefficient and hard to understand.
New Dictionary: You have fewer words, but each word is a perfect shape. One word describes "the circle," another describes "the sphere."
The Result: Their new tool reconstructed the AI's thoughts much more accurately (less error) than the old tools. It found that the AI's brain is full of these complex, multi-dimensional shapes, not just simple lines.

5. The "Ghost" in the Machine (Stability)

One interesting finding is that even if you train the tool twice, it might find different specific "shapes" (different dictionaries), but the overall room they describe is the same.

Analogy: Imagine two different artists painting the same landscape. Artist A uses blue and green; Artist B uses purple and orange. They use different colors (different dictionary entries), but they both paint the same mountain and river (the same underlying structure). The authors proved that while the specific "colors" change, the "landscape" remains stable.

Summary

The paper claims that AI models are not just collections of straight lines. They are filled with curved, multi-dimensional shapes. By using a new tool (Bilinear Autoencoders) that can see these curves, researchers can:

See more clearly: They find concepts that were previously invisible or messy.
Be more efficient: They need fewer "features" to describe the same amount of information.
Understand better: They can see how concepts like "years" or "locations" are actually structured as geometric shapes, not just simple switches.

The authors have even built an interactive website (a visualizer) where you can look at these 3D shapes yourself, proving that these curved "manifolds" are real and prevalent in how AI thinks.

Technical Summary: Bilinear Autoencoders Find Interpretable Manifolds

Problem Statement

Current mechanistic interpretability relies heavily on Sparse Autoencoders (SAEs) to uncover interpretable latent representations in neural networks. These methods operate under the Linear Representation Hypothesis, assuming that concepts are encoded along linear directions. However, recent research indicates that salient concepts often span complex, multi-dimensional manifolds (e.g., ellipsoids, hyperboloids) that linear methods cannot capture without post-hoc analysis. Furthermore, the isolated study of linear latents often fails to account for context-dependence, where a latent's meaning shifts based on co-active features. While increasing model capacity can capture more structure, unbounded expressivity often leads to trivial solutions that defeat interpretability. The core challenge is to find a representation class that is nonlinear enough to capture these manifolds but algebraically tractable enough to allow for closed-form geometric analysis and compositional understanding.

Methodology

The paper proposes Bilinear Autoencoders as a solution, extending the standard sparse autoencoder architecture to learn quadratic latents.

Architecture

Instead of reconstructing the input vector $x$ directly, the bilinear autoencoder reconstructs the lifted product space $X = xx^\top$ .

Encoder: The model uses Gated Linear Units (GLUs) adapted to compute second-order polynomials. The activation of the $i$ -th latent is defined as $z_i = x^\top W_i x$ , where $W_i$ is a symmetric matrix representing a quadratic form.
Decoder: The reconstruction is performed linearly in the lifted space: $\hat{X} = W^\top W X$ .
Training Efficiency: Directly materializing the $d \times d$ matrix $X$ is computationally intractable for large models. The authors employ a kernel trick to compute the reconstruction loss and gradients without explicitly constructing $X$ . The kernel matrix $K = WW^\top$ is computed efficiently via the factorization $K = C^\top (LL^\top \odot RR^\top) C$ , where $L$ and $R$ are encoder factor matrices and $C$ is a composition matrix.
Sparsity: The authors introduce Hoyer density as a scale-invariant sparsity penalty. Unlike standard $L_1$ penalties that can cause trivial shrinkage, Hoyer density encourages relative prominence across the batch, ensuring latents fire selectively on specific structures without collapsing to zero.

Structural Hypotheses

The paper formalizes three nested representation hypotheses that define the geometric priors of the dictionary atoms:

Quadratic Representation Hypothesis: The dictionary consists of symmetric matrices $W_i$ . The model decomposes $xx^\top$ into a sparse sum of these matrices. This allows for the detection of quadric geometries (ellipsoids, hyperboloids, paraboloids).
Atomic Representation Hypothesis: A constraint where each dictionary atom is a rank-1 matrix ( $W_i = w_i w_i^\top$ ). The receptive field is a symmetric "slab" (two parallel hyperplanes), detecting energy along a direction regardless of sign.
Composite Representation Hypothesis: Dictionary atoms are sparse mixtures of rank-1 matrices ( $W_i = \sum C_{ij} w_j w_j^\top$ ). This allows the model to learn low-rank subspaces automatically. A composite atom can form an ellipsoid (if positive semi-definite) or a hyperboloid (if indefinite), effectively capturing concepts that require multiple linear directions to define.

Key Contributions

Architecture: The authors adapt bilinear autoencoders for dictionary learning, enabling efficient training via a kernel trick. This provides a natural, tractable extension of linear SAEs that operates in weight space, allowing for geometric analysis without forward passes.
Framework: They propose a formal framework of structural priors (Atomic, Composite, Quadratic) that map dictionary constraints to specific receptive field geometries (slabs, ellipsoids/hyperboloids, and arbitrary quadrics). This challenges the standard linear hypothesis by demonstrating that concepts can be detected via their proximity to quadric surfaces rather than linear half-spaces.
Empirical Validation:
- Reconstruction: Experiments on Qwen 3.5 (0.8B), Gemma-3 (1B), and Llama-3.2 (1B) show that composite and quadratic latents consistently achieve lower reconstruction error (NMSE) than atomic (rank-1) latents and linear baselines.
- Geometry Prevalence: Spectral analysis reveals that while many concepts are low-rank, a significant "long tail" of concepts requires higher-dimensional geometries (rank > 1) to be captured accurately.
- Stability: While individual dictionary atoms (decompositions) vary significantly across training runs and hyperparameters, the global input subspace recovered by the autoencoders remains highly stable (>90% similarity). This suggests that the specific decomposition is less important than the underlying manifold structure.
- Qualitative Analysis: An interactive visualizer demonstrates that bilinear autoencoders capture diverse geometric patterns, including cones (years), clusters (locations), interpolations (semantic bridges), and subsets (contextual distinctions like "for" vs. "for instance").

Significance and Claims

The paper argues that bilinear autoencoders represent a significant step toward nonlinear but mathematically tractable latent representations.

Beyond Linearity: The work challenges the dominance of the linear representation hypothesis, showing that salient concepts often inhabit multi-dimensional manifolds that linear methods miss.
Compositional Interpretability: By using polynomials, the model ensures that composition happens in weight space. This allows for explicit, computable interactions between features, moving interpretability from isolated bases to understandable compositional structures.
Manifold Discovery: The authors position these models as unsupervised tools for manifold discovery. Because the geometry is read directly from the weight tensors (eigenvalues and eigenvectors of $W_i$ ), the models offer a "lens" on representation geometry that is interpretable by design.
Modest Scope: The authors explicitly state that the structural priors presented are a proof-of-concept. They do not claim to have swept all possible priors to find the optimal trade-off, nor do they claim to have solved causal ablation. Instead, they frame this as a foundation for future work in causal steering and characterizing which geometric priors specific models prefer.

The paper concludes that while linear methods are useful, the field must move toward representations that can explicitly model the curved, high-dimensional geometry inherent in neural network activations to achieve true compositional understanding.