Tunable Wavelet Units (UwUs) Overview

Updated 7 July 2026

Tunable Wavelet Units (UwUs) are parameterized wavelet constructions that expose key design variables to enable flexible adaptations across reverse holography, CNN downsampling, and unitary-circuit designs.
They replace conventional pooling and strided convolutions in neural networks by integrating learnable wavelet filter banks that improve image classification, anomaly detection, and segmentation performance.
Empirical results show that UwUs provide measurable gains in accuracy and detail preservation, with performance improvements reported on benchmarks ranging from CIFAR-10 to 3D OCT segmentation.

Tunable Wavelet Units (UwUs) denote a set of wavelet-based constructions in which a small number of parameters are exposed as explicit design or learning variables. In the cited literature, the term covers several non-identical objects: a one-parameter family of affine-coherent wavelets used in reverse holography and emergent geometry (Vedl et al., 9 Apr 2025); trainable two-channel wavelet filter-bank modules that replace pooling, stride-two convolution, or downsampling in convolutional networks (Le et al., 1 Jul 2025); orthogonal, biorthogonal, and lifting-based variants for image classification, anomaly detection, and volumetric OCT segmentation (Le et al., 21 Jul 2025, Le et al., 1 Jul 2025, Le et al., 22 Jul 2025); and closely related formulations in unitary-circuit wavelet design, filterbank autoencoders, scale-translation equivariant networks, beta-derived compactly supported wavelets, and warped filter-bank frames (Evenbly et al., 2016, Jawali et al., 2021, Romero et al., 2020, Oliveira et al., 2015, Holighaus et al., 2014). The unifying feature is tunability of the wavelet basis or filter bank, rather than a single canonical mathematical definition.

1. Terminological scope and major usages

The term is used across multiple research programs, and the underlying objects differ substantially in domain, parameterization, and purpose. In reverse holography, UwUs are wavelets in continuous scale space whose parameter $\sigma$ tunes the emergent bulk curvature (Vedl et al., 9 Apr 2025). In CNNs, UwUs are end-to-end trainable building blocks that insert a small two-channel, orthogonal or biorthogonal, perfect-reconstruction wavelet filter bank in place of conventional max-pooling, strided convolution, or naive down-sampling (Le et al., 21 Jul 2025). In unitary-circuit and wavelet-design literatures, closely related objects are parameterized by local rotations, lifting coefficients, regularizers, or warping functions (Evenbly et al., 2016, Jawali et al., 2021, Holighaus et al., 2014).

Context	Construction	Tunable parameters
Reverse holography	Continuous wavelet basis on a field $\phi(x)$	scale $a$ , width $\sigma$
CNN downsampling	OrthLatt-UwU / PR-Relax-UwU	$\theta_k$ , $h_0(n)$ , $\alpha$
3D OCT segmentation	OrthLatt-UwU / BiorthLatt-UwU / LS-BiorthLattUwU	$\theta_k$ , $k_m$ , $a_k$
Unitary-circuit design	Dyadic WT as layers of $\phi(x)$ 0 rotations	$\phi(x)$ 1
Learned filterbank design	Conv-autoencoder wavelet learning	support, moments, symmetry weights
Continuous compact-support families	Beta-derived unicycle wavelets	$\phi(x)$ 2
Warped filter-bank frames	Non-uniform filter banks	$\phi(x)$ 3, $\phi(x)$ 4, $\phi(x)$ 5

A common misconception is to treat UwU as the name of one fixed architecture or one fixed mother wavelet. The literature instead uses the term for several families of tunable wavelet constructions. This suggests that “UwU” functions more as a design pattern—explicitly parameterized wavelet analysis embedded in a larger computational pipeline—than as a single standardized transform.

2. Continuous-wavelet UwUs and emergent geometry

In "Emergent metric from wavelet-transformed quantum field theory" (Vedl et al., 9 Apr 2025), the construction is purely spatial. In one spatial dimension, the continuous wavelet transform is

$\phi(x)$ 6

with daughter wavelet

$\phi(x)$ 7

In momentum space,

$\phi(x)$ 8

The admissibility condition on the mother wavelet is

$\phi(x)$ 9

which guarantees invertibility.

The paper introduces a one-parameter family of wavelets based on affine-group coherent states. In the log-momentum variable $a$ 0, the pair $a$ 1 with $a$ 2 is canonical, so coherent states are Gaussians in $a$ 3. Transforming back to $a$ 4 yields a mother wavelet in momentum space whose tunable parameters are the usual wavelet scale $a$ 5 and the Gaussian standard deviation $a$ 6. The paper states

$a$ 7

so $a$ 8 controls the joint uncertainty $a$ 9 (Vedl et al., 9 Apr 2025).

The central application is reverse holography. Using the Petz-Rényi mutual information $\sigma$ 0 between two infinitesimally close wavelet modes $\sigma$ 1 and $\sigma$ 2, the emergent $\sigma$ 3-dimensional metric takes the Poincaré-half plane form

$\sigma$ 4

where $\sigma$ 5 is a fixed reference scale and the curvature radius $\sigma$ 6 depends only on the wavelet’s moments. For free massless fermionic and bosonic QFTs, the emerging metric is asymptotically anti-de Sitter space, and the geometry is tunable by the chosen wavelet basis (Vedl et al., 9 Apr 2025).

The practical interpretation is unusually direct. The scale coordinate $\sigma$ 7 plays the role of the emergent bulk radial coordinate, while the single parameter $\sigma$ 8 sets the bulk radius of curvature through $\sigma$ 9. The paper states the limiting behavior

$\theta_k$ 0

and recommends choosing a target curvature radius $\theta_k$ 1, solving $\theta_k$ 2, fixing $\theta_k$ 3, building the mother wavelet in momentum space, generating wavelets at all scales via $\theta_k$ 4, transforming the boundary QFT into the $\theta_k$ 5 picture, and extracting the metric locally from $\theta_k$ 6 (Vedl et al., 9 Apr 2025). The same source notes that, in the coincidence limit $\theta_k$ 7, wavelet overlaps make two-point correlators ill-defined, whereas using $\theta_k$ 8 yields a UV-finite, basis-independent distance measure.

3. Discrete filter-bank parameterizations

In CNN-oriented papers, UwUs are typically discrete two-channel filter banks with learnable low-pass and high-pass filters. A central branch is the Orthogonal Lattice–based Wavelet Unit, or OrthLatt-UwU. In the $\theta_k$ 9-domain, the analysis filters are parameterized by a cascade of Givens-rotation matrices and delays,

$h_0(n)$ 0

where

$h_0(n)$ 1

and the filter order is $h_0(n)$ 2 (Le et al., 1 Jul 2025). Because each $h_0(n)$ 3 is orthonormal and $h_0(n)$ 4 only delays one branch, the composite filter bank is guaranteed to be orthogonal and therefore admits perfect reconstruction with no additional constraints.

The Perfect-Reconstruction-Relaxation Wavelet Unit, or PR-Relax-UwU, keeps the two classical perfect-reconstruction constraints but enforces them softly. Alias cancellation is imposed by

$h_0(n)$ 5

while the half-band condition is penalized through

$h_0(n)$ 6

and the total loss is

$h_0(n)$ 7

This construction trades strict perfect reconstruction for greater adaptability, while maintaining the alias-cancellation relation between low and high channels (Le et al., 1 Jul 2025).

A second branch relaxes orthogonality more radically by moving to biorthogonal lifting. In "Biorthogonal Tunable Wavelet Unit with Lifting Scheme in Convolutional Neural Network" (Le et al., 1 Jul 2025), the polyphase representation is factorized into lifting stages

$h_0(n)$ 8

with

$h_0(n)$ 9

The only free parameters are the scalars $\alpha$ 0, and the lifting construction automatically guarantees perfect reconstruction as a biorthogonal system.

A further extension adds a stop-band penalty to orthogonal lattice units. In "Stop-band Energy Constraint for Orthogonal Tunable Wavelet Units in Convolutional Neural Networks for Computer Vision problems" (Le et al., 21 Jul 2025), the stop-band energy of $\alpha$ 1 is

$\alpha$ 2

with numerical approximation over $\alpha$ 3 sample points, and the normalized stop-band loss is

$\alpha$ 4

Training uses

$\alpha$ 5

The stated objective is to make $\alpha$ 6 behave like a true low-pass filter and $\alpha$ 7 like a high-pass filter.

Across these variants, the main design degrees of freedom are rotation angles $\alpha$ 8, direct filter coefficients $\alpha$ 9, lifting scalars $\theta_k$ 0, lattice scalars $\theta_k$ 1, or loss weights such as $\theta_k$ 2. Theoretical guarantees vary accordingly: exact orthogonality and perfect reconstruction may hold by construction, or they may be relaxed through soft penalties in order to increase discriminative flexibility.

4. Integration into neural architectures

The canonical CNN use of a UwU is to replace resolution-reducing operators with wavelet analysis followed by subband fusion. In the ResNet18 backbone used for ERM surgery classification, the initial $\theta_k$ 3 stride-2 convolution, the downsample $\theta_k$ 4 convolution in the residual block, and the $\theta_k$ 5 max-pool are each replaced by a non-strided convolution followed by a wavelet decomposition and one shallow fusion layer (Le et al., 1 Jul 2025). If $\theta_k$ 6 is the incoming feature map and $\theta_k$ 7 denote the learned low-pass and high-pass analysis matrices, the four subbands are

$\theta_k$ 8

A single $\theta_k$ 9 convolution then combines the activated subbands: $k_m$ 0

The same high-level pattern appears in the stop-band-constrained orthogonal units. Each UwU replaces a pooling or downsampling layer; instead of discarding high-frequency detail, it performs a learned $k_m$ 1 critically sampled wavelet analysis, applies nonlinearities to each subband, and then learns an optimal linear fusion back into a half-resolution feature map (Le et al., 21 Jul 2025). For stride-2 convolutions, the original stride-2 layer is replaced by stride-1 convolution followed by a UwU. For pure spatial downsampling, an analysis-only UwU can pass only the $k_m$ 2 subband followed by a $k_m$ 3 fusion layer.

In volumetric retinal segmentation, the same idea is transferred to 3D OCT data. "Universal Wavelet Units in 3D Retinal Layer Segmentation" replaces each max-pool in a motion-corrected MGU-Net with a UwU down-sampling block, and the decoder uses the corresponding synthesis bank (Le et al., 22 Jul 2025). At each encoding stage, the pipeline is: $k_m$ 4 convolution, batch normalization, and ReLU; two-branch wavelet analysis into four subbands; a subband attention head with learnable channel-wise weights; and fusion to the next stage. The decoder mirrors this with synthesis, refinement convolution, and skip connections.

The computational cost is modest but nonzero. For one UwU block with $k_m$ 5 input channels and $k_m$ 6-tap filters, the analysis stage contributes approximately $k_m$ 7 parameters, and the $k_m$ 8 fusion from $k_m$ 9 contributes $a_k$ 0 parameters (Le et al., 1 Jul 2025). On an ImageNet-sized ResNet18, replacing all three downsampling stages with orthogonal lattice units was reported as roughly $a_k$ 1– $a_k$ 2 more parameters and $a_k$ 3– $a_k$ 4 more FLOPs, with inference slowdown under $a_k$ 5 and peak memory increase of only about $a_k$ 6 (Le et al., 21 Jul 2025). A related lifting-based implementation reported a practical overhead of $a_k$ 7– $a_k$ 8 on top of a standard ResNet forward pass (Le et al., 1 Jul 2025).

5. Reported empirical performance

The medical-imaging and vision papers report consistent gains over corresponding baselines, though the gains depend strongly on task and variant. In ERM surgery classification from postoperative OCT center scans, the baseline ResNet18 on original scans reached $a_k$ 9 accuracy, preprocessing with energy crop and wavelet denoising reached $\phi(x)$ 00, OrthLatt-UwU reached $\phi(x)$ 01, and PR-Relax-UwU reached $\phi(x)$ 02 (Le et al., 1 Jul 2025). The same paper reports that a trained human grader achieved $\phi(x)$ 03 accuracy on the postoperative OCT classification task.

Setting	Baseline	Best reported UwU result
ERM surgery classification	$\phi(x)$ 04	$\phi(x)$ 05 with PR-Relax-UwU
CIFAR-10, ResNet-18	$\phi(x)$ 06	$\phi(x)$ 07 with SBE-OrthLatt-UwU
DTD, ResNet-18	$\phi(x)$ 08	$\phi(x)$ 09 with SBE unit
MVTec hazelnut, Det-AUROC	$\phi(x)$ 10	$\phi(x)$ 11 with SBE-UwU
JRC OCT segmentation, Avg. Dice	$\phi(x)$ 12	$\phi(x)$ 13 with LS-BiorthLatt-UwU

The stop-band-constrained orthogonal units show especially large gains on texture-rich data. On CIFAR-10 with ResNet-18, the baseline was $\phi(x)$ 14 and SBE-OrthLatt-UwU reached $\phi(x)$ 15, a gain of $\phi(x)$ 16 (Le et al., 21 Jul 2025). On the Describable Textures Dataset, the same baseline of $\phi(x)$ 17 rose to $\phi(x)$ 18, a gain of $\phi(x)$ 19. In ResNet-34, the DTD accuracy increased from $\phi(x)$ 20 to $\phi(x)$ 21 with SBE-UwU. On MVTec hazelnut anomaly detection, using a DTD-trained SBE-UwU ResNet-18 as the CFLOW-AD encoder yielded Seg-AUROC $\phi(x)$ 22 versus $\phi(x)$ 23 baseline and Det-AUROC $\phi(x)$ 24 versus $\phi(x)$ 25 baseline (Le et al., 21 Jul 2025).

The lifting-based biorthogonal units report similar trends. In ResNet-18 image classification, the baseline was $\phi(x)$ 26 on CIFAR-10 and $\phi(x)$ 27 on DTD; OrthLatt-UwU-2Tap reached $\phi(x)$ 28 and $\phi(x)$ 29; LS-BiorUwU-2Step reached $\phi(x)$ 30 and $\phi(x)$ 31; and LS-BiorUwU-3Step reached $\phi(x)$ 32 and $\phi(x)$ 33 (Le et al., 1 Jul 2025). The same paper reports a clear task dependence: increasing lifting steps yielded diminishing returns on CIFAR-10 but clearer gains on DTD, which it characterizes as rich in high-frequency textures.

In 3D retinal layer segmentation from OCT volumes, the MGU-Net baseline with max-pool achieved average Dice $\phi(x)$ 34 and average voxel-wise accuracy $\phi(x)$ 35. OrthLatt-UwU reached $\phi(x)$ 36, BiorthLatt-UwU reached $\phi(x)$ 37, and LS-BiorthLatt-UwU reached $\phi(x)$ 38 (Le et al., 22 Jul 2025). The paper identifies LS-BiorthLatt-UwU as the best trade-off between parameter efficiency and high-frequency preservation on that dataset.

These results support a narrow empirical claim: replacing fixed downsampling with trainable wavelet filter banks can improve classification, anomaly detection, and segmentation when the target task depends on fine detail, texture, or structural consistency. A stronger claim about general superiority across all architectures or data regimes would go beyond the cited evidence.

Several adjacent literatures help clarify what is structurally distinctive about UwUs. In "Representation and design of wavelets using unitary circuits" (Evenbly et al., 2016), a length- $\phi(x)$ 39 dyadic wavelet transform is written as an orthogonal matrix

$\phi(x)$ 40

where each layer is a direct sum of nearest-neighbor $\phi(x)$ 41 rotations

$\phi(x)$ 42

This provides a minimal parametrization for orthogonal wavelets and a factorization algorithm for known filter banks such as Daubechies, coiflets, and symlets. The same framework extends to symmetric dilation-3 wavelets, multi-wavelets, boundary wavelets, and biorthogonal wavelets.

In "Wavelet Design in a Learning Framework" (Jawali et al., 2021), a single-level wavelet transform plus inverse is cast as a purely linear two-layer convolutional autoencoder. Analysis uses stride-2 convolutions with $\phi(x)$ 43 and $\phi(x)$ 44, synthesis upsamples and convolves with dual filters, and training minimizes

$\phi(x)$ 45

The paper states that near-zero loss implies perfect reconstruction with very high probability. Orthogonality, compact support, smoothness, symmetry, and vanishing moments are incorporated through architecture design and regularization, and the approach recovers known Daubechies and Cohen-Daubechies-Feauveau families while also learning wavelets outside those families.

In "Wavelet Networks: Scale-Translation Equivariant Learning From Raw Time-Series" (Romero et al., 2020), each UwU implements a learnable, scale-translation equivariant wavelet transform followed by point-wise nonlinearity. The lifting convolution

$\phi(x)$ 46

generalizes the continuous wavelet transform on a dyadic scale grid, and a subsequent group convolution mixes information along the scale-translation group. The paper states that these are exactly the most general linear maps equivariant to continuous translation and scaling.

Continuous wavelet families offer additional forms of tunability. Beta-derived compactly supported one-cycle wavelets are parameterized by $\phi(x)$ 47 through a Beta density on $\phi(x)$ 48 or an affinely rescaled interval $\phi(x)$ 49, with

$\phi(x)$ 50

and the limit $\phi(x)$ 51 recovers Haar (Oliveira et al., 2015). Warped filter-bank frames parameterize non-uniform frequency tilings through a warping function $\phi(x)$ 52 and decimation factors $\phi(x)$ 53, with warped responses

$\phi(x)$ 54

and tightness characterized by

$\phi(x)$ 55

(Holighaus et al., 2014).

Open directions are already explicit in the 2025 CNN literature. Reported proposals include multi-level decompositions within a single layer, extension to 3D or spatiotemporal data, learning layer- or channel-specific stop-band cutoffs $\phi(x)$ 56, incorporating paraunitary wavelet constraints into attention heads of Vision Transformers, and exploring unsupervised or self-supervised pretraining with wavelet consistency losses (Le et al., 21 Jul 2025). This suggests that the research trajectory is moving from isolated wavelet layers toward a broader program in which multiresolution structure, perfect reconstruction, equivariance, and task-driven tunability are treated as co-optimizable architectural primitives.