A guided journey · 32 lessons
Sequence models & research
Build the foundations for attention, rotary position, state-space memory and reproducible research claims.
Prerequisites are included automatically. Suggested lesson time: 17 hours, plus extra time for experimentation and projects. This is an estimate, not a measured completion time.
Your route
Work down the list, or open any lesson. Progress is shared across paths: a lesson you have completed elsewhere counts here too. Completing checks is a useful milestone, not proof of mastery.
Progress loads from this browser when JavaScript is available.
- What is a signal?What does a waveform actually tell you?
- Frequency, amplitude and phaseHow can three numbers describe an oscillation?
- Vectors and similarityHow does a list of numbers become a geometric object?
- Complex numbers and rotationWhy do transforms use an imaginary coordinate?
- Slopes and gradientsHow do you decide which way to change a parameter?
- Probability and uncertaintyHow should several plausible outcomes share one unit of probability?
- Data splits and leakageWhen is a good test score too good to be true?
- Sampling and timeHow do equally spaced measurements encode motion?
- Aliasing and NyquistCan different waves produce exactly the same samples?
- Projection onto wavesHow can a dot product detect a frequency?
- The discrete Fourier transformWhat exactly is in an FFT output array?
- Windows and leakageWhy does one tone make many spectral bars?
- Time-frequency analysisWhat if the signal’s frequency changes over time?
- Linearity and time invarianceWhen can one impulse response describe a whole system?
- Convolution step by stepWhich input samples contribute to one output?
- Frequency response and filter trade-offsWhy does averaging suppress some frequencies more than others?
- Recurrence and stabilityWhen does yesterday’s state fade, persist or explode?
- Features and linear regressionWhat exactly does a model learn from examples?
- Gradient descentHow can a repeated local update find better parameters?
- Classification and cross-entropyWhy penalise a confidently wrong prediction so strongly?
- Nonlinear features and neural networksWhat does adding a hidden layer actually add?
- The chain rule and backpropagationHow does one weight receive credit for a distant error?
- Position and representationHow can a model distinguish the same content in different places?
- Attention as weighted retrievalHow can a query choose which information to combine?
- Rotary positionWhy does a shared position shift disappear from a rotary score?
- State-space memoryWhen is a recurrence also a convolution?
- Nonlinear aliasingCan a cleanly sampled signal create aliases later?
- FFTs and memory trafficWhy can fewer arithmetic operations still run slower?
- Numerical stabilityCan the same exact formula behave differently on a machine?
- Benchmarking and reproducibilityWhat evidence supports a claim that an implementation is faster?
- Project: a sequence-model investigationWhich sequence mechanism suits the dependency you need?
- Project: read and reproduce a research claimWhat would convince you that a result is both correct and useful?
After a lesson, return here to pick the next step in this path. With JavaScript enabled, lesson next/previous links stay in this route. Without JavaScript, they follow the complete course; return here to choose your next route lesson. Export or import your shared notebook on the course page.