⚡ electrical engineering

FUTON: Fourier Tensor Network for Implicit Neural Representations

The paper introduces FUTON, a novel Implicit Neural Representation that models signals using a low-rank tensor decomposition of Fourier basis functions to achieve faster training, better convergence, and superior performance in image representation and inverse problems compared to state-of-the-art MLP-based approaches.

Original authors: Pooya Ashtari, Pourya Behmandpoor, Nikos Deligiannis, Aleksandra Pizurica

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Pooya Ashtari, Pourya Behmandpoor, Nikos Deligiannis, Aleksandra Pizurica

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 describe a beautiful, complex painting to a friend over the phone.

The Old Way (MLPs):
Most current AI methods (called MLP-based INRs) try to describe the painting by listing every single brushstroke, one by one, in a long, winding sentence. They say, "Start at the top left, go down a bit, make a blue dot, then a red dot, then a green smear..."

The Problem: This takes forever to say (slow convergence). If you get distracted and say the wrong color once, the whole picture gets messed up (overfitting to noise). And if you try to guess what the painting looks like outside the frame, you have no idea because you only memorized the specific strokes you were told (poor extrapolation).

The New Way (FUTON):
The paper introduces FUTON (Fourier Tensor Network). Instead of listing every single dot, FUTON looks at the painting and says, "Ah, this is mostly a big blue wave, with some red ripples on top, and a few tiny green sparkles."

It breaks the image down into waves (Fourier series) and then realizes that these waves aren't random; they follow a simple, organized pattern (Low-Rank Tensor).

Here is how FUTON works, using some everyday analogies:

1. The "Musical Score" Analogy (Fourier Series)

Imagine the image isn't a picture, but a song.

Old AI tries to write out the song by listing every single note played by every instrument at every millisecond. It's a massive, messy list.
FUTON writes a musical score. It identifies the main melody (smooth waves) and the harmony (periodic patterns). It knows that a song is just a combination of a few simple notes played together. This makes it much faster to understand and recreate the song.

2. The "Packing a Suitcase" Analogy (Low-Rank Tensor)

Now, imagine you have to pack all the ingredients for a massive banquet into a suitcase.

The Old Way: You pack every single grain of rice, every drop of sauce, and every spice individually. The suitcase is huge, heavy, and takes forever to pack.
FUTON: It realizes that the rice, the sauce, and the spices all come from the same few "flavor profiles." It packs them into compact, pre-mixed flavor packets.
- Instead of packing 10,000 individual items, it packs 50 "flavor packets" that can be mixed in different ways to recreate the whole banquet.
- This is the Low-Rank part. It compresses the data by finding the hidden patterns that connect everything.

3. Why is FUTON Better?

Speed (The Express Train): Because FUTON uses these "flavor packets" (mathematical shortcuts) instead of listing every single item, it trains 2 to 5 times faster than the old methods. It's like taking a high-speed train instead of walking through every street in the city.
Less Mistakes (Anti-Noise): If you give FUTON a slightly blurry photo (noisy data), it doesn't try to memorize the blur. Because it understands the "big picture" waves, it naturally ignores the static and fills in the gaps correctly. It's like listening to a song on a bad radio; you still hear the melody clearly, even if there's static.
Better Guessing (Extrapolation): If you ask FUTON to draw a part of the image it hasn't seen yet, it can guess correctly because it understands the rules of the waves, not just the specific dots. It knows that if the wave goes up here, it should go down there.

The "Secret Sauce": The Tensor Network

The paper uses a fancy math trick called a Tensor Network.
Think of this like a Lego set.

The old AI tries to build a castle by gluing together millions of tiny, unique Lego bricks.
FUTON realizes that the castle is built from just a few types of standard Lego bricks, arranged in a specific, repeating pattern. It only needs to remember the pattern and the types of bricks, not every single connection. This saves a massive amount of memory and computing power.

Real-World Results

The authors tested FUTON on:

Reconstructing Images: It made clearer, sharper pictures than the competition.
Fixing Blurry Photos (Super-Resolution): It could turn a tiny, pixelated image into a high-definition one better than anyone else.
Removing Noise (Denoising): It cleaned up grainy, noisy photos without making them look blurry or plastic.
Medical Scans (CT): It could reconstruct clear images of lungs and bones from very few X-ray angles, which is crucial for reducing radiation exposure for patients.

In a Nutshell

FUTON is a smarter, faster, and more efficient way for computers to "remember" and "recreate" images and 3D objects. Instead of memorizing every single pixel like a robot, it learns the underlying rhythm and structure of the data, allowing it to work faster, make fewer mistakes, and create higher-quality results.

1. Problem Statement

Implicit Neural Representations (INRs), or neural fields, have become a dominant paradigm for encoding signals (images, volumes, radiance fields) as continuous functions parameterized by neural networks. However, standard Multi-Layer Perceptron (MLP) based INRs suffer from three critical limitations:

Slow Convergence: Training often requires extensive time to converge.
Overfitting: They are prone to overfitting noise when regularization is insufficient, particularly in inverse problems.
Poor Extrapolation: They struggle to generalize to unobserved regions, which is crucial for tasks like super-resolution and reconstruction.

These issues stem from a lack of inductive biases in MLPs that align with the low-dimensional spectral structure inherent in natural signals. While specialized activations (e.g., SIREN, WIRE) improve performance, they still rely on deep, non-linear mappings that scale poorly and lack explicit structural priors.

2. Methodology: FUTON

The authors propose FUTON (Fourier Tensor Network), a novel INR architecture that models signals as generalized Fourier series where the coefficients are parameterized via a low-rank tensor decomposition.

Core Architecture

Instead of learning arbitrary non-linear mappings via deep MLPs, FUTON expresses a signal $s(x)$ as a weighted combination of orthonormal, separable basis functions:
$\hat{s}_K(x) = \sum_{k \in \mathcal{K}} w^{(k)} \phi_k(x)$
Where:

$\phi_k(x)$ are separable basis functions (e.g., products of 1D cosine functions).
$w^{(k)}$ are the Fourier coefficients.

Low-Rank Parameterization (CP Decomposition)

The key innovation is how the weight tensor $W$ (containing all coefficients $w^{(k)}$ ) is parameterized. A full-rank tensor $W \in \mathbb{R}^{K \times \dots \times K \times D}$ would require exponential parameters ( $O(K^C D)$ ). FUTON employs Canonical Polyadic (CP) decomposition to factorize $W$ into a sum of $R$ rank-1 tensors:
$W = \sum_{r=1}^R u^{(1)}_r \otimes u^{(2)}_r \otimes \dots \otimes u^{(C)}_r \otimes v_r$
This results in a compact set of learnable factor matrices $U^{(c)}$ and $V$ .

Computational Efficiency

By substituting the CP decomposition into the Fourier series, the forward pass is restructured into a sequence of efficient operations:

Projection: Project 1D basis vectors onto factor matrices $U^{(c)}$ (Matrix-Vector multiplication).
Combination: Combine projected features via Hadamard (element-wise) products.
Output Mapping: Map the result to the output dimension via matrix multiplication with $V$ .

Complexity: This approach reduces the computational complexity from exponential $O(K^C D)$ to linear $O(CKR + CR + DR)$ with respect to input dimension $C$ and output dimension $D$ .

Inductive Biases

FUTON combines two complementary biases:

Fourier Bases: Capture smoothness and periodicity, naturally handling high-frequency details.
Low-Rank Factorization: Enforces a low-dimensional spectral structure, acting as an intrinsic regularizer that prevents overfitting and aids generalization.

3. Key Contributions

Novel Architecture: Introduction of FUTON, which unifies generalized Fourier series with low-rank tensor decomposition, offering a structured alternative to MLP-based INRs.
Theoretical Guarantees: Proof of a Universal Approximation Theorem, demonstrating that FUTON can approximate any $L_2$ function to arbitrary precision given sufficient basis resolution ( $K$ ) and rank ( $R$ ).
Computational Efficiency: Derivation of an inference algorithm with linear complexity in spectral resolution and input dimension, enabling scalability to high-dimensional signals.
Superior Performance: Extensive empirical validation showing FUTON outperforms state-of-the-art MLPs (SIREN, WIRE, Gauss) in reconstruction quality, convergence speed, and generalization on inverse problems.

4. Experimental Results

The authors evaluated FUTON on signal representation and inverse problems using the Kodak dataset (images), Thai Statue (volumes), DIV2K (super-resolution/denoising), and medical CT scans.

Signal Representation

Image Reconstruction: On the Kodak dataset, FUTON achieved 37.12 dB PSNR and 96.54% SSIM, significantly outperforming the best baseline (WIRE: 33.69 dB PSNR).
Convergence Speed: FUTON trained 2–5× faster than WIRE and other MLP-based methods, reaching high-quality reconstructions in seconds.
3D Volumes: On occupancy volume representation, FUTON achieved 99.67% IoU, surpassing WIRE (99.08%) with smoother surfaces and fewer artifacts.

Inverse Problems

Super-Resolution: FUTON recovered high-frequency details better than bilinear interpolation and MLPs, achieving 30.74 dB PSNR (vs. 30.54 dB for WIRE).
Image Denoising: In the presence of sensor noise, FUTON achieved 27.04 dB PSNR, effectively removing noise while preserving edges, outperforming WIRE by 0.53 dB.
CT Reconstruction: In sparse-view CT reconstruction, FUTON produced clearer anatomical structures with reduced streak artifacts, achieving 29.18 dB PSNR (vs. 27.80 dB for WIRE).

Ablation Studies

Rank ( $R$ ): Performance saturates around $R=512$ , confirming the low-rank nature of natural signal spectra.
Basis Functions: The Cosine basis significantly outperformed Legendre and Chebyshev polynomials in both accuracy (37.12 dB vs. ~33 dB) and convergence speed, likely due to the sparsity of natural images in the frequency domain and efficient hardware evaluation.

5. Significance

FUTON represents a paradigm shift in Implicit Neural Representations by moving away from "black box" deep MLPs toward structured, interpretable, and mathematically grounded models.

Efficiency: It solves the scalability bottleneck of Fourier-based methods via tensor networks, making high-resolution signal representation feasible.
Generalization: The low-rank spectral parameterization acts as a powerful implicit prior, making FUTON exceptionally robust for inverse problems where data is sparse or noisy.
Practicality: With faster training times and higher reconstruction quality, FUTON offers a practical, drop-in replacement for MLP-based INRs in applications ranging from computer vision to medical imaging.

In summary, FUTON demonstrates that explicitly modeling the spectral structure of signals via low-rank tensor decomposition yields superior performance, speed, and generalization compared to learning arbitrary non-linear mappings with deep neural networks.