Scattering Transforms

Scattering transforms offer a rare, transparent view into how signal processing works before any training happens. There is no learning from data, no guessing, and the behavior is guaranteed in advance. This makes it the clearest possible entry point for understanding how convolutional networks actually process information. By mastering this, you gain a solid foundation for exploring modern AI systems, including large language models, which rely on similar principles of pattern recognition and stability.

Imagine you have a one-dimensional signal—like a sound wave or a stock price over time—and you want to understand its structure. A scattering transform does this by repeatedly applying two operations: a filter that looks for specific local patterns (such as a sudden jump or a smooth curve), and an averaging operation that summarizes a local region by taking the average of values within it.

Crucially, these filters are fixed. They are not learned from data but are designed using base functions used to generate these filters to be sensitive to different frequencies and scales. The process builds a hierarchy: first it looks at the signal itself, then it captures the energy of the filtered patterns at different scales, then it sees how that energy varies across those scales, and so on.

For example, if you take a segment of audio, calculate its energy in one frequency band, then look at how that energy changes over time using another band, you get a description of the signal's structure. This output is stable to small deformations (like a slight time shift) and invariant to translations (shifting the whole signal does not change the result). Because everything is deterministic and linear—except for taking the absolute value of numbers (which we call modulus), which introduces non-linearity—the properties can be proven mathematically. You know in advance how the transform will behave, unlike a standard neural network where you only hope it generalizes well.

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 base functions used to generate these filters—where one stencil detects sharp edges, another detects gentle slopes, and another detects periodic patterns. 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. This gives a summary of what patterns are present and roughly where. But that is 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 centimeters, 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 core idea of fixed templates, local matching, and summarizing is exactly right.

It works like a fingerprint scanner that doesn't care where your finger lands on the pad, only that the ridges match the expected pattern.

The scattering transform is a fixed architecture that applies a cascade of wavelet filters interleaved with modulus (taking the absolute value) 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) and Pooling as a Statistic (averaging is the pooling here, and it serves as a sufficiency argument). It also connects closely to Receptive Fields (deeper scattering layers see wider context) and Fourier Analysis of Networks (scattering transforms are often analyzed via their frequency response).