Scattering Transforms

Scattering transforms give you a convolutional network that works before any training—no learned weights, no backpropagation—and its behavior is guaranteed in advance. This makes it the clearest possible introduction to how convolutional networks actually process signals, and it is the main analytical tool for understanding why they work. Mastering this unlocks deeper study of convolutional layers, pooling, receptive fields, and Fourier analysis of networks—topics you already have notes on—and gives you a solid foundation for exploring how modern AI systems, including large language models, rely on similar principles.

Imagine you have a one-dimensional signal—like a sound wave or a stock price over time—and you want to understand it. A scattering transform works by repeatedly applying two operations: a wavelet filter, which is like a template that looks for a specific local pattern (e.g., a sudden jump or a smooth curve), and a pooling operation, which summarizes a local region by taking the average. Crucially, these filters are fixed—they are not learned from data. They are designed analytically (using mother wavelets) to be sensitive to different frequencies and scales. The transform builds a hierarchy: first it averages the signal itself, then it captures the energy of wavelet coefficients at different scales, then it captures how that energy varies across scales, and so on. The output is a set of coefficients that describe the signal's structure in a way that is stable to small deformations (like a slight time shift or a minor shape change) and invariant to translations (shifting the whole signal does not change the result). Because everything is deterministic and linear (except for the modulus, which is a pointwise absolute value), the properties can be proven mathematically—you know in advance how the transform will behave, unlike a trained neural network where you only hope it generalizes.

Think of a scattering transform as a team of inspectors examining a manuscript. The manuscript is the signal. The team has a set of fixed stencils (the wavelets)—one for detecting sharp edges, one for detecting gentle slopes, one for detecting periodic patterns, and so on. Each inspector slides their stencil across the page and notes how well it matches at each position. The raw match values are then pooled: for each stencil, the team records the average absolute match over a window (pooling). This gives a summary of what patterns are present and roughly where. But that's not enough—the team also wants to know how the pattern changes as you move your finger along the page. So they take another set of stencils, these ones finer, and slide them across the sequence of pooled matches, recording how much the match fluctuates. They pool that too. By iterating this procedure—stencil, match, take absolute value, pool—they build a hierarchical description that is robust to small shifts: if you shift the entire manuscript by a few centimetres, the team's summary barely changes, because they are looking at the pattern of matches, not the exact positions. This is exactly the translation invariance and deformation stability of the scattering transform. The analogy breaks down when you consider that real wavelets are continuous functions—the stencils are not binary shapes but smoothly varying curves—and the modulus (absolute value) operation is not something a human inspector would naturally do. But the core idea of fixed templates, local matching, and summarizing is exactly right.

The scattering transform is a fixed convolutional architecture that applies a cascade of wavelet filters interleaved with modulus and averaging operators, producing coefficients that are translation invariant and stable to small deformations, with properties that can be proven before any training.

This concept is the analytical backbone of convolutional networks. It sits at the intersection of your notes on Convolutional Layers (where filters are learned, here they are fixed), Pooling as a Statistic (averaging is the pooling here, and it's a sufficiency argument), and Receptive Fields (deeper scattering layers see wider context). Fourier Analysis of Networks is a direct neighbour—scattering transforms are often analysed via their frequency response. It also builds on the idea of weight sharing from Convolutional Layers, but here the weights are not learned—they are chosen analytically. You have not yet studied wavelets in detail, but this concept will introduce them gently.