SCRAPL: Scattering Transform with Random Paths for Machine Learning
Christopher Mitcheltree, Vincent Lostanlen, Emmanouil Benetos, Mathieu Lagrange
TL;DR
SCRAPL introduces a stochastic approach to efficiently compute multivariable scattering transforms, enabling their use in neural network training for perceptual quality assessment and audio signal processing tasks.
Contribution
The paper proposes SCRAPL, a novel stochastic optimization scheme that reduces computational costs of scattering transforms, facilitating their integration into deep learning workflows.
Findings
SCRAPL significantly speeds up scattering transform computations.
Applying SCRAPL improves neural network convergence in audio tasks.
SCRAPL enhances perceptual quality assessment in deep inverse problems.
Abstract
The Euclidean distance between wavelet scattering transform coefficients (known as paths) provides informative gradients for perceptual quality assessment of deep inverse problems in computer vision, speech, and audio processing. However, these transforms are computationally expensive when employed as differentiable loss functions for stochastic gradient descent due to their numerous paths, which significantly limits their use in neural network training. Against this problem, we propose "Scattering transform with Random Paths for machine Learning" (SCRAPL): a stochastic optimization scheme for efficient evaluation of multivariable scattering transforms. We implement SCRAPL for the joint time-frequency scattering transform (JTFS) which demodulates spectrotemporal patterns at multiple scales and rates, allowing a fine characterization of intermittent auditory textures. We apply SCRAPL to…
Peer Reviews
Decision·ICLR 2026 Poster
• SCRAPL makes the Joint Time-Frequency Scattering (JTFS) transform feasible for use in the loss function of large-scale deep learning by overcoming its prohibitively high computational and memory costs. • The paper proposes an innovative stochastic approximation scheme—sampling just a single path—which guarantees an unbiased estimate of the true loss gradient in expectation, together with a novel Path-Wise Optimization (P-Adam) to manage the high variance of the single-path sampling, stabilizin
• Limited Experimental Scope and Generalizability: The evaluation is performed on a relatively limited number of sound matching examples, which does not permit concluding on the general applicability of the method across different audio domains or tasks. • When listening to the audio examples, one should note that while the sound matching algorithm converges, the perceptual similarity of the original and generated sounds is not generally satisfying. In quite a few of the examples, the output sou
SCRAPL is a novel and technically well-motivated approach that makes scattering transforms practical for end-to-end differentiable learning. The paper is clearly written and grounded in both theory and application. The proposed stochastic approximation is theoretically justified (unbiased gradient under uniform sampling) and supported by well-designed optimization variants (P-Adam, P-SAGA) that address variance and non-i.i.d. gradient issues. The θ-importance sampling heuristic is a creative con
While the proposed framework is promising, certain limitations remain. The experiments are largely confined to audio applications, and it is unclear how well SCRAPL generalizes to image or other modalities that use scattering transforms. Although θ-importance sampling improves convergence, its computation involves complex gradient–Hessian interactions that may limit scalability in higher-dimensional settings. The empirical comparisons, while comprehensive, rely on relatively small neural network
* A stochastic approach that makes optimization under scattering transforms computationally efficient * Empirical results demonstrating that proposed approach achieves accuracies within a factor of two relative to the extensive joint time-frequency scattering transforms, while being within 3x of the much more efficient multi-scale spectral loss approach. * Clear, well written paper
* The empirical results on a sound matching setup which is of relatively limited interest to the broader NeurIPS and ML community. Results on more audio generation tasks utilizing perceptual qualities of scattering transforms will strengthen the paper significantly.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsSparse and Compressive Sensing Techniques · Speech and Audio Processing · Image and Signal Denoising Methods