C.W.K.
Stream
Lesson 02 of 05 · published

The S4 → S4D → H3 → S6 Lineage

~14 min · s4, hippo, h3, history

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

S4 made SSMs practical with HiPPO

S4 (Gu, Goel, Ré, 2022) was the breakthrough that turned SSMs from a control-theory curiosity into a credible deep-learning architecture. The trick was the HiPPO initialization (High-order Polynomial Projection Operator), which gives the A matrix a structured form that's mathematically optimal for compressing the input history into a polynomial basis. With HiPPO, the model could remember information from thousands of timesteps back without exploding or collapsing.

S4 also worked out the convolution view of the linear recurrence: when A is fixed, the SSM can be "unrolled" into a global convolution kernel K = (CB, CAB, CA²B, CA³B, ...). Applying that kernel via FFT gives O(n log n) training, fully parallel on GPUs. Same model, two views: recurrent at inference, convolutional at training. S4 dominated the Long Range Arena benchmark and proved that SSMs could be more than a footnote.

S4D and H3 — simplification and gating

S4D simplified S4 by restricting A to be diagonal. The full HiPPO matrix had a complex structured form that was awkward to implement; the diagonal restriction lost a small amount of expressiveness for a much friendlier GPU implementation. This was the engineering pattern that would repeat: simplify the math to win on the hardware, recover quality with bigger models or other tricks.

H3 (Hungry Hungry Hippos, 2023) added gating mechanisms to SSM blocks — multiplicative interactions inspired by the gates in LSTMs and GRUs. With H3, SSMs started approaching Transformer quality on language modeling for the first time, especially on associative-recall-style tasks. But all of S4, S4D, and H3 were still Linear Time-Invariant.

The LTI ceiling

LTI means the matrices A, B, C are fixed, regardless of input. The same A is applied at every timestep — no matter what the current token is. This is what enables the convolution view (a global kernel exists because A doesn't depend on x), but it's also a hard ceiling on expressiveness. An LTI SSM cannot decide "this particular token is important — remember it more" or "this token is filler — fade it out". Every input is processed by the same time-invariant filter.

The S6 step — "Selective State Space", which became Mamba — broke LTI by making B, C, and Δ functions of the input. That's the next track. Everything in this lesson is the runway that selectivity finally took off from.

External links

Exercise

Read Sasha Rush's Annotated S4 (linked above) end-to-end. You don't need to fully internalize the HiPPO derivation — but pay attention to the moment the paper unrolls the recurrence into a convolution kernel. That "two views of the same model" trick is the engineering soul of every SSM up through pre-Mamba designs, and understanding it is what makes Mamba's choice to give it up legible.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.