02
Designing
CRF Transition Matrix Explorer
When does a sequence prior outweigh strong token-level evidence?
Matrix perturbation · path sensitivity
Computational notebook · Lab 01—06
Inspect the decision, not only the output.
Small, inspectable instruments for reasoning about structured prediction, retrieval, and video understanding.
1 runnable study5 instrument notes
Current instrument / 01
A decoding trace you can interrupt.
Adjust the local evidence, then watch dynamic programming preserve only the best partial sequence for each label. The last step follows those saved decisions backward.
Interactive notebook · 01
Viterbi decoder
Find the highest-scoring label sequence without enumerating every possible path.
Step 0 / 4Set the evidence
xAspirint1relievest2paint3
A · local evidence
Emission scores
Strong local evidence makes the best path easy to read.
| Token | B-DRUG | O | B-SYM |
|---|---|---|---|
| Aspirin | |||
| relieves | |||
| pain |
B · sequence prior
Transitions
| Aᵢⱼ | B-DRUG | O | B-SYM |
|---|---|---|---|
| B-DRUG | -0.6 | 0.8 | 0.1 |
| O | 0.2 | 0.5 | 0.7 |
| B-SYM | -0.5 | 0.4 | -0.2 |
C · dynamic program
Score table + backpointers
δt(j) = et(j) + maxi[δt−1(i) + Aij]| State | Aspirint1 | relievest2 | paint3 |
|---|---|---|---|
| B-DRUGdrug | — | — | — |
| Ooutside | — | — | — |
| B-SYMsymptom | — | — | — |
Set the evidence
Choose a preset or edit an emission. Then advance to calculate one token at a time.
Focus this panel: use ←/→ to step, space to play, and R to reset.
How to read it
Scores accumulate left to right.
Each table cell combines one emission with the best incoming transition.
Arrows preserve argmax choices.
They are backpointers, not attention weights or probabilities.
Backtracking is delayed.
The optimal labels become known only after the final column is scored.
Notebook queue / 02—06
Instruments being shaped
Each entry begins with one technical question and one inspectable artifact. These are design notes, not finished demos.
03
Planned
Biomedical Entity Linking Retrieval Explorer
How do lexical and dense retrieval fail on different clinical mentions?
Candidate ranks · score decomposition
04
Prototyping
Shot Boundary Timeline
Which temporal signals separate a hard cut from rapid camera motion?
Frame strip · boundary confidence
05
Planned
Keyframe Similarity Explorer
How much visual redundancy remains after shot-aware sampling?
Embedding distance · threshold sweep
06
Research note
Attention Map Viewer
What changes when attention is compared across heads and layers?
Token matrix · head comparison