← Back to the interactive worksheet
Reusable lecture assets / SVG

Make the mechanism visible.

Six original figures, with readable numbers and explicit assumptions. Save each SVG for slides, or use the worksheet's frame-download button to export another animation beat. All figures work offline.

A film is more than its frames.

Three numbered binary frames. Reverse their order to reverse motion; the time-averaged image stays unchanged.
Three numbered binary frames. Reverse their order to reverse motion; the time-averaged image stays unchanged.

Same numerical model and renderer as the live worksheet.

Download SVG ↓

What should come out—and when?

One-hot inputs feed recurrent states. This exported frame shows one review label after the final state; the worksheet also shows token tags and shifted next-token targets.
One-hot inputs feed recurrent states. This exported frame shows one review label after the final state; the worksheet also shows token tags and shifted next-token targets.

Same numerical model and renderer as the live worksheet.

Download SVG ↓

Read. Combine. Update.

The third tanh update for inputs [1,0,1], input weight 1, recurrent weight 0.5, bias 0 and initial state 0. Values displayed to three decimals.
The third tanh update for inputs [1,0,1], input weight 1, recurrent weight 0.5, bias 0 and initial state 0. Values displayed to three decimals.

Same numerical model and renderer as the live worksheet.

Download SVG ↓

One cell. Many uses.

Four uses of one parameter bank. With input width 3, hidden width 2 and output width 2 there are 18 parameters, including biases.
Four uses of one parameter bank. With input width 3, hidden width 2 and output width 2 there are 18 parameters, including biases.

Same numerical model and renderer as the live worksheet.

Download SVG ↓

Carry one clue through the gap.

A clue passes through 20 zero-input linear updates with recurrent weight 0.8. The final state is approximately 0.011529; this is not an accuracy.
A clue passes through 20 zero-input linear updates with recurrent weight 0.8. The final state is approximately 0.011529; this is not an accuracy.

Same numerical model and renderer as the live worksheet.

Download SVG ↓

Can the correction reach the beginning?

Backward sensitivity through 20 linear links of derivative 0.8. The vertical axis is logarithmic. The footer supplies separate scalar parameter-gradient clipping examples; the plotted path is never clipped.
Backward sensitivity through 20 linear links of derivative 0.8. The vertical axis is logarithmic. The footer supplies separate scalar parameter-gradient clipping examples; the plotted path is never clipped.

Same numerical model and renderer as the live worksheet.

Download SVG ↓

Printable student questions · Instructor notes and answers