Parametric Scattering Networks

• Parametric scattering networks are a family of architectures that integrate tunable wavelet parameters with scattering transforms to produce adaptable, invariant representations.
• They employ parameterized filterbanks, such as Morlet wavelets and spectral filters on graphs, optimized via gradient-based methods for enhanced performance.
• The models maintain theoretical stability and computational efficiency, demonstrating improved accuracy in low-data regimes and large-scale simulations.

Parametric scattering networks constitute a family of architectures that leverage mathematical properties of scattering transforms while introducing explicit tunable parameters, either learned or physically defined, into the design of wavelet, filterbank, or junction components. These models generalize the original fixed-parameter scattering transform to enable data-driven or application-driven adaptation while preserving core invariance and stability guarantees. Research on parametric scattering networks spans Euclidean, graph, physical-science, and acoustic domains, with corresponding parameterization strategies, computational pipelines, and theoretical results.

1. Mathematical Foundations and Formulation

Parametric scattering transforms are based on the cascade of convolutional filters, modulus nonlinearities, and local averaging, but with filterbanks parameterized and potentially optimized rather than fixed a priori.

Given an input signal $x \in \mathbb{R}^n$ (for Euclidean or image settings), the classic $m$-th order scattering coefficient along path $p = (\lambda_1, ..., \lambda_m)$ is defined by

$$U[p]x(u) = \big|\cdots \big| |x * \psi_{\lambda_1}| * \psi_{\lambda_2} \big| \cdots * \psi_{\lambda_m} \big| (u)$$

with final local averaging by a scaling function $\phi_J$: $S[p]x(u) = (U[p]x * \phi_J)(2^J u)$ where each $\psi_{\lambda}$ is a (typically complex) wavelet parameterized by scale $j$ and orientation $\ell$.

Parametric scattering replaces the standard fixed collection $\{\psi_{\lambda}\}_{\lambda}$ with a set of wavelets $m$0 parameterized by vectors $m$1—for instance, the Morlet family: $m$2 where $m$3 (scale), $m$4 (orientation), $m$5 (frequency), and $m$6 (aspect ratio) are parameters subject to optimization or physical specification (Gauthier et al., 2021). In graph scenarios, spectral filters $m$7 with parametric $m$8 families (polynomial, rational, etc.) are employed (Koke et al., 2023).

The transform's core contraction and invariance properties derive from its combination of unitary (or tight-frame) operators and $m$9-type modulus nonlinearity, with guaranteed stability to small diffeomorphisms (Mallat et al., 2013).

2. Parametric Filterbank Construction and Learning

In traditional wavelet scattering, filterbank parameters (such as the mother wavelet's scale, frequency, orientation array) are fixed to form a tight frame or orthonormal system. In parametric scattering networks, these parameters become learned or explicitly tuned components:

• Euclidean/Signal Domains: The canonical Morlet wavelet parameters $p = (\lambda_1, ..., \lambda_m)$0 may be learned via supervised or unsupervised objectives. Learning proceeds by backpropagation through differentiable filter generators (with provided gradients in [(Gauthier et al., 2021), Appendix D]) and passes gradients through FFT-based convolutions.
• Graph Domains: Filters $p = (\lambda_1, ..., \lambda_m)$1 are constructed by applying a parameterized function (e.g., polynomials, rational functions, Fourier multipliers) of the graph Laplacian $p = (\lambda_1, ..., \lambda_m)$2. The coefficients of these functions act as learnable or cross-validated parameters (Koke et al., 2023).
• Physical Networks: Parameters correspond to physical quantities, such as permittivity distributions in electromagnetic scattering, delay lines and scattering matrices in acoustic networks, or wave propagation constants specified by physical laws (Sena et al., 2015, Valantinas et al., 2022).

Empirical studies (Gauthier et al., 2021) show that learning the four canonical Morlet parameters for each wavelet significantly improves performance in small-sample regimes compared to fixed (tight-frame) designs, especially for object/medical classification tasks.

3. Training and Objective Functions

In supervised settings, parametric scattering networks optimize filter parameters $p = (\lambda_1, ..., \lambda_m)$3 jointly with downstream classifier weights via minimization of regularized cross-entropy: $p = (\lambda_1, ..., \lambda_m)$4 using SGD with momentum, weight decay, and a learning rate schedule (Gauthier et al., 2021). For representations, either a linear classifier or a hybrid convolutional block (e.g., Wide-ResNet) may follow scattering.

For unsupervised pretraining, contrastive losses (SimCLR) can be used to adapt $p = (\lambda_1, ..., \lambda_m)$5 via augmentations (Gauthier et al., 2021). In the classical unsupervised scenario, the filterbank remains fixed and unsupervised learning optimizes contraction objectives to preserve variance while increasing discriminative spread (Mallat et al., 2013).

Physical-science parametric networks (e.g., for Maxwell’s equations) are "trained" by physically specifying the kernel weights and running the network to convergence—no SGD, so the solution is deterministic (Valantinas et al., 2022).

4. Implementation Details and Computational Properties

Implementations in image domains utilize frameworks such as Kymatio (with custom differentiable parameterizations for Morlet filters) (Gauthier et al., 2021). For graph domains, functional calculus is applied to the Laplacian, enabling efficient matrix operations and guaranteed stability (Koke et al., 2023). In physical networks, FFT-based convolutions, pointwise multiplications, and preconditioned iterative solvers (e.g., Richardson iteration for Maxwell) underpin efficient computation (Valantinas et al., 2022). Acoustic parametric scattering networks (Scattering Delay Networks, SDNs) are constructed as networks of delay lines and lossless scattering matrices, with parameters derived from geometric and absorption properties of the simulated enclosure (Sena et al., 2015).

Computational complexity depends on parameterization:

• For scattering on large grids, cost is dominated by FFTs ($p = (\lambda_1, ..., \lambda_m)$6 per recurrence in physics-defined nets (Valantinas et al., 2022)).
• In graph settings, the branching ratio and number of paths exponentially inflate feature dimensionality, but functional calculus filtering and Lipschitz nonlinearities ensure numerically tractable and stable propagation (Koke et al., 2023).
• In SDNs, cost is dominated by the number of nodes ($p = (\lambda_1, ..., \lambda_m)$7) and per-node filter length, scaling as $p = (\lambda_1, ..., \lambda_m)$8 FLOPs per sample (Sena et al., 2015).

Memory usage in physics-defined scattering networks scales linearly with grid points, with reported requirements of up to 35 GiB for hundreds of millions of nodes (Valantinas et al., 2022).

5. Empirical Results and Quantitative Comparisons

Parametric scattering networks show consistent improvement over fixed-parameter scattering and match or outperform learned convolutional baselines in limited-data and structured-regime tasks:

• On CIFAR-10, learnable scattering + linear classifier (LS+LL) achieves up to +4.56% accuracy improvement over tight-frame scattering for 500 training examples; with WRN backends, the gain persists (+1.98% for 1K samples) (Gauthier et al., 2021).
• In COVIDx-CRX2 (medical X-ray images), learnable scattering delivers +1.22–1.45% accuracy boost in extreme low-data settings (Gauthier et al., 2021).
• For acoustic SDNs, the parametric design achieves first-order reflection accuracy matching the Image Method, while requiring orders-of-magnitude fewer computations and negligible memory ($p = (\lambda_1, ..., \lambda_m)$9200 kB) (Sena et al., 2015).
• In graph domains, flexible parametric graph-scattering architectures outperform classical wavelet graph scattering and standard graph-based learning approaches in both classification and regression tasks (Koke et al., 2023).
• In physics-defined scattering networks for Maxwell’s equations, computation of light-scattering in millimeter-scale volumes is tractable with massive grid sizes (up to 576 million points), with deterministic, non-biased solutions and no training cost (Valantinas et al., 2022).

6. Theoretical Properties and Stability Guarantees

Core invariance and stability results are preserved under parameterization, provided key constraints (tight frames, Lipschitz nonlinearity) are maintained:

• Scattering networks with learned wavelet parameters maintain stability to small diffeomorphisms and translation (Mallat et al., 2013, Gauthier et al., 2021).
• Graph-scattering frameworks admit explicit, spectrally agnostic stability guarantees for both node- and graph-level perturbations—including stability under vertex-set changes—through bounds on frame constants and Lipschitz parameters (Koke et al., 2023).
• Theoretical results confirm that tight-frame initialization provides strong inductive bias in low-data regimes, while the learned parameters yield more discriminative, dataset-adaptive invariants without sacrificing deformation or perturbation stability (Gauthier et al., 2021, Koke et al., 2023).
• In physical networks, deterministic convergence to a physically correct solution is guaranteed up to discretization and convergence tolerances (Valantinas et al., 2022).

Energy propagation, feature truncation, and aggregation strategies carry over to parametric designs: both the Euclidean and graph cases provide energy decay guarantees, bounding feature stability as network depth $$U[p]x(u) = \big|\cdots \big| |x * \psi_{\lambda_1}| * \psi_{\lambda_2} \big| \cdots * \psi_{\lambda_m} \big| (u)$$0 increases (Gauthier et al., 2021, Koke et al., 2023).

7. Extensions, Applications, and Open Directions

Parametric scattering networks generalize to numerous signal domains:

• Images and signals: Adaptive wavelet bank learning for image/textural/spectral analysis and small-sample classification.
• Graphs: Node and graph-invariant representations for topological learning, quantum chemistry, and social network analysis (Koke et al., 2023).
• Physical sciences: Electromagnetic (Maxwell) and acoustic (SDN) wave propagation, with direct mapping from physical parameters to network weights and efficient simulation (Valantinas et al., 2022, Sena et al., 2015).
• Hybrid architectures: Integration of parametric scattering front-ends with trainable deep-network back-ends for enhanced performance in large-data or complex settings (Gauthier et al., 2021).

Outstanding challenges include automating end-to-end learning of all parametric components (log-transform offsets, scattering matrices), extending invariance to broader transformation groups (e.g., rotation, scale), and optimizing networks for ultra-large-scale and on-device deployment (Singh et al., 2017, Gauthier et al., 2021).

Parametric scattering networks thus provide a mathematically principled, computationally efficient, and highly flexible approach to invariant feature extraction, with strong theoretical guarantees and demonstrated empirical success across modalities (Mallat et al., 2013, Gauthier et al., 2021, Koke et al., 2023, Valantinas et al., 2022, Sena et al., 2015, Singh et al., 2017).