⚡ electrical engineering

Mechanistic Interpretability with Sparse Autoencoder Neural Operators

This paper introduces Sparse Autoencoder Neural Operators (SAE-NOs), a novel framework that extends traditional sparse autoencoders by parameterizing concepts as structured functions rather than scalar activations, thereby enabling the modeling of concept expression across input domains, achieving superior generalization across resolutions, and accelerating optimization through a novel lifting technique.

Original authors: Bahareh Tolooshams, Ailsa Shen, Anima Anandkumar

Published 2026-05-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Bahareh Tolooshams, Ailsa Shen, Anima Anandkumar

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 complex machine works by looking at its internal gears. In the world of Artificial Intelligence (AI), specifically in "mechanistic interpretability," researchers use tools called Sparse Autoencoders (SAEs) to find these gears, which they call "concepts."

Traditionally, these tools have been a bit like a simple checklist. If a concept is present, they give it a single number (a "scalar") to say how strong it is. For example, "The concept of 'cat' is active with a strength of 0.8."

This paper introduces a new, more powerful tool called SAE-NOs (Sparse Autoencoder Neural Operators), and specifically a version called SAE-FNOs. Here is the simple breakdown of what they did and why it matters, using everyday analogies.

1. The Old Way: The "On/Off" Switch vs. The "Map"

The Problem:
Think of a standard SAE (SAE-MLP) like a light switch for a room. It can tell you if the light is on or off, and maybe how bright it is. But it can't tell you where in the room the light is shining or how the light is moving. If you have a picture of a cat walking across a room, the old system might say, "Okay, 'cat' is on," but then it has to turn that switch off and turn on a different "cat" switch when the cat moves to the next spot. It treats every position as a completely new thing.

The New Solution:
The new SAE-FNO is like a dynamic map or a spotlight. Instead of just a switch, it describes the concept as a function that can move and change shape. It can say, "The concept of 'cat' is active, and here is exactly where it is in the image and how it is shaped."

Analogy: Imagine you are describing a song.
- Old Way: You just say, "The song is playing."
- New Way: You describe the melody, the rhythm, and how the notes move across time. You capture the structure of the song, not just its existence.

2. Two Types of "Sparsity" (Being Selective)

The paper introduces a clever way to be selective in two different ways at once:

Concept Sparsity (The "Who"): This decides which concepts are active. (e.g., "Only the 'cat' and 'tree' concepts are on; turn off the 'car' concept.")
Domain Sparsity (The "Where"): This decides where those active concepts appear. (e.g., "The 'cat' concept is only active in the top-left corner of the image, not the whole image.")

The Magic: By combining these, the system can reuse the same "cat" concept over and over again, just moving it to different spots on the map. The old system had to invent a new "cat" for every new spot. This makes the new system much more efficient, like using one reusable sticker and moving it around, rather than printing a new sticker for every single spot.

3. The "Fourier" Secret Sauce

To make this work, the researchers used something called Fourier Neural Operators.

The Metaphor: Imagine you are trying to describe a complex wave in the ocean. You could try to describe every single drop of water (which is hard and messy). Or, you could describe the wave by its frequency and shape (like a smooth, rolling curve).
The Benefit: The new system describes concepts as smooth waves (functions) rather than jagged pixels. This allows it to understand patterns that are smooth or structured, like edges in a photo or waves in a sound, much better than the old "pixel-by-pixel" approach.

4. Key Superpowers Discovered

The paper tested this new system on images (like MNIST digits and CIFAR photos) and found several cool things:

Stability: If you change how "strict" the system is about being selective (sparsity), the old system's concepts get messy and change completely. The new system keeps its concepts stable and consistent, no matter the settings.
Moving Objects: When a digit (like a '3') moves across a screen, the old system switches between dozens of different "concept IDs" to track it. The new system keeps the same concept ID and just shifts its location on the map. It understands that it's the same object moving, not a series of different objects.
Zooming In and Out (Generalization): This is a big one. If you train the old system on small images (28x28 pixels), it breaks if you show it a bigger image (56x56). It's like learning to drive on a small parking lot and then failing on a highway. The new system, because it learns "functions" (rules) rather than fixed grids, can handle bigger images or different resolutions it has never seen before. It generalizes like a human who understands the concept of a car, not just the specific pixels of one car.
Lifting (The Preconditioner): The paper also added a step called "lifting," which temporarily boosts the data into a higher-dimensional space to make learning easier, then brings it back down. They proved mathematically that this acts like a preconditioner (a tool that smooths out a bumpy road), helping the AI learn faster and more accurately.

Summary

In short, this paper says: "Stop treating AI concepts like static light switches. Start treating them like moving, shape-shifting functions."

By moving from simple numbers to complex functions (using Fourier math), the new system:

Learns where and how concepts appear, not just if they appear.
Reuses concepts efficiently (saving memory and computation).
Stays stable even when you change the rules.
Can handle images of different sizes without breaking.

The authors claim this is a fundamental shift in how we build tools to understand AI, moving from rigid, fixed-grid thinking to flexible, functional thinking.

Technical Summary: Sparse Autoencoder Neural Operators (SAE-NOs)

1. Problem Statement

Standard Sparse Autoencoders (SAEs) have become a primary tool for mechanistic interpretability, aiming to uncover latent concepts in neural networks by enforcing sparsity on activations. However, existing formulations (referred to as SAE-MLPs) rely on the Linear Representation Hypothesis (LRH), which models data as sparse linear combinations of scalar-activated concepts. This approach imposes three critical limitations:

Scalar Activations: Concepts are represented by a single scalar value, encoding only presence and strength, but failing to capture how or where a concept is expressed across the input domain.
Fixed Discretization: Concepts are parameterized as vectors in the same Euclidean space as the input, tying the representation to a specific resolution or grid. This prevents generalization to different discretizations or resolutions.
Inductive Bias: For spatially structured data (e.g., images, physical fields), standard SAEs force concepts into globally uniform contributions, limiting their ability to represent localized or domain-dependent structures. While Convolutional SAEs (SAE-CNNs) introduce locality, they remain tied to fixed receptive fields and resolutions.

The paper argues that moving from vector-valued to functional parameterizations is necessary to model structured concept expression, particularly for data exhibiting spatial or temporal structure across scales.

2. Methodology: SAE-NOs and SAE-FNOs

The authors introduce Sparse Autoencoder Neural Operators (SAE-NOs), a framework where concepts are parameterized as functions rather than vectors. This is instantiated specifically as Sparse Autoencoder Fourier Neural Operators (SAE-FNOs).

Functional Representation Hypothesis (FRH)

The authors formalize the Functional Representation Hypothesis (FRH), positing that data $x$ in a function space $\mathcal{X}$ can be explained by a sparse representation $z$ in a function space $\mathcal{Z}$ via a linear operator $G_{D^*}$ :
$x = G_{D^*}(z)$
Unlike standard SAEs where $z$ is a vector of scalars, here $z$ is a sample from a function space, and the "dictionary" $D$ is an operator.

Joint Sparsity Mechanism

SAE-FNOs introduce a dual-sparsity structure to encode not just which concepts are active, but how and where:

Concept Sparsity (Group Sparsity): Selects a small subset of active concepts (groups).
Domain Sparsity (Intra-group Sparsity): Constrains the activation pattern within the feature map of each active concept. This determines the spatial location of the concept's expression.

Mathematically, for a concept $c$ , the activation is a feature map $v_c$ . Concept sparsity selects $c$ based on the $\ell_2$ -norm of $v_c$ , while domain sparsity applies element-wise sparsity within $v_c$ .

Fourier Parameterization

SAE-FNOs parameterize the concepts as integral operators in the Fourier domain. The decoder reconstructs the input as:
$\hat{x} = \mathcal{F}^{-1} \left( \sum_{c=1}^p W_c \cdot (\mathcal{F} z_c) \right)$
where $W_c$ are learnable Fourier coefficients (kernels) and $z_c$ are the sparse feature maps. This formulation induces an inductive bias toward sparse convolutional representations.

Lifting

The framework incorporates lifting, a linear operator $L$ that maps inputs to a higher-dimensional space before concept learning, followed by a projection $P$ back to the original space. Theoretically, the authors show that lifting acts as a preconditioner, rescaling update directions to accelerate optimization and improve dictionary recovery.

3. Key Contributions

SAE-NO Framework: Introduction of a new class of sparse autoencoders operating in function spaces, enabling resolution-invariant representations.
Functional Sparse Coding: Replacement of scalar activations with feature-map representations, decoupling concept expressivity (controlled by concept sparsity) from structural locality (controlled by domain sparsity).
Theoretical Characterization:
- Proof that lifting acts as a preconditioner, accelerating convergence.
- Demonstration that under specific conditions (orthogonal lifting, tied projection), lifted SAE-NOs reduce to standard SAE-NOs, ensuring architectural equivalence.
- Characterization of training dynamics showing that SAE-FNOs recover concepts under a sparse convolutional generative model.
Emergent Temporal Consistency: Demonstration that SAE-FNOs naturally reuse concepts across translated patterns (e.g., moving digits) by shifting spatial activations, without requiring explicit temporal constraints, sequence-level training, or temporal memory.

4. Experimental Results

The authors evaluate SAE-FNOs against SAE-MLPs and SAE-CNNs on synthetic data, MNIST, CIFAR-10, and ImageNet.

Concept Recovery: On synthetic data with known ground-truth convolutional dictionaries, SAE-FNOs successfully recover underlying concepts, outperforming SAE-CNNs, especially under large initialization errors.
Stability Across Sparsity: SAE-FNOs learn concepts whose structural characteristics (measured by Hoyer score) remain stable as sparsity levels ( $k$ ) vary. In contrast, SAE-MLP concepts change drastically and lose structure as $k$ increases.
Efficiency and Expressivity: SAE-FNOs achieve higher reconstruction accuracy ( $R^2$ ) with fewer active concepts. They exhibit a sharper decay in concept usage, indicating more efficient reuse of structured concepts compared to the flatter distribution of SAE-MLPs.
Generalization Across Discretization: SAE-FNOs trained at a specific resolution (e.g., 28x28) generalize to higher resolutions (e.g., 56x56, 112x112) with minimal error increase. Standard SAEs fail to generalize beyond their training discretization, showing sharp increases in reconstruction error.
Robustness: SAE-FNOs exhibit greater stability under high-frequency noise perturbations and spectral shifts compared to SAE-MLPs.
Correlated Concepts: Unlike SAE-MLPs which tend toward orthogonal features, SAE-FNOs learn a broader range of correlated concepts, as the functional parameterization allows concepts to differ in where they are expressed, reducing interference.

5. Significance and Claims

The paper claims that the shift from vector-valued to functional parameterization fundamentally extends the capabilities of sparse autoencoders. By moving beyond the "presence" of a concept to modeling "structured concept expression," SAE-FNOs offer:

Resolution Invariance: The ability to operate on data at resolutions unseen during training.
Structural Stability: Concepts that remain consistent across varying sparsity levels and domain sizes.
Efficient Representation: A mechanism to represent data populations with fewer effective concepts by reusing them across spatial domains.

The authors position SAE-NOs as a general functional framework rather than a specific architecture, noting that the FNO parameterization is orthogonal to the sparsification mechanism. They conclude that functional parameterization is central to learning representations for structured domains, highlighting the importance of how concepts are parameterized in mechanistic interpretability.

Limitations Acknowledged:
The paper notes that SAE-FNOs incur higher computational and memory costs due to Fourier operations and feature-map representations. Additionally, the spectral bias toward smooth, low-frequency patterns may be less suitable for highly irregular, discontinuous, or tabular data where vector-based SAEs might remain preferable. The empirical study is focused on vision datasets, leaving applications to neural operator foundation models and scientific machine learning as future directions.