05Family 03 of 05

The engine

Evidence on one frame is noisy. The engine turns 25 opinions a second into one coherent story of the half, and the ranking metric scores the story. Three decisions, taken in order.

Your question The study that answers it Where should time live? Evidence, then a decoder Which decoder, and why? Seven ways to decode a match Several models disagree? Experts, gates and referees
The decision map of this family. Three questions, in the order they have to be answered.
  • 3studies in this family
  • 7decoders compared on the same scores
  • 2levels at which experts can be combined

What moves through the engine

The engine changes the unit of the problem three times: from opinions on frames, to scores, to states, to possessions. The ranking metric reads only the last one, which is why a small error near the top can cost a whole possession at the bottom.

Thirty frames in, two possessions out. The tie-break metric reads the states. The ranking metric reads the possessions.

The trade-off surface of the family

Every option in this family buys capacity with labels and pays for it in what it can promise about the sequence. Put them on one plane and the constraint from the previous family, the labels that exist, draws a line through it.

Across: the labels an engine needs. Up: the guarantee it can state about the sequence. With no labels on the match, only the left column can be fielded. Positions are judgements, not measurements.

Three decisions, in order

Decision The options What decides it
Where time lives Inside one multi-frame model, or in a decoder behind the evidence The labels that exist; what must be auditable
Which decoder Majority, Viterbi, HMM, CRF, HSMM, change-point, neural What it needs, what it guarantees, how it treats short spells
Who decides when models disagree A gate on every frame, or a referee on a disputed segment Whether the experts’ errors are independent

The second question only exists if the first was answered “a stage of its own”. The third only exists if there is more than one expert or more than one decoded sequence.

How an architect chooses here

  • Ask what labels exist. A learner is only as good as its labels. A decoder with fixed costs needs none.
  • Prefer a guarantee you can state. “The best path under these two costs” can be argued with. A learned smoothness cannot.
  • Compare on the spell, not the frame. The metric counts possessions, so the engine is judged on possessions.
  • Earn every added part. An expert whose removal changes nothing was a passenger.

What this family cannot do

No engine repairs evidence that was never there. A missing ball, a swapped identity and a stoppage nobody tagged are problems of representation, and they belong to the hard parts.

In this family3

  • Where should temporal consistency live: inside one model that reads many frames, or in a stage of its own behind the evidence? One learner, or two machines in a chain. The answe...

    • One model or two
    • Decoder
    • Auditable
    • Fielded
  • The same per-frame scores, seven different stories. What each decoder assumes about football, what it needs, what it guarantees and how it fails.

    • Viterbi
    • HMM
    • CRF
    • HSMM
  • When no single model is right everywhere, several can be combined: frame by frame through a gate that knows whom to trust, or segment by segment through a referee that rules on ...

    • Ensemble
    • Gating
    • Adjudication