Part IV · Ch. 20 — HMM (Hidden Markov Model)
Probabilistic · Sequences · Markov chain + emissions. Models sequences with latent states that emit observations. Card:
../02-types-of-ai/03-probabilistic.kmd.
🎨 Figure
F-IV.20.0— States that hide. Brief: chain of hidden states (circles) linked over time, each emitting an observation (squares); Viterbi path highlighted.
1. Definition and short history
Models a Markov process whose states are not observed directly. Dominated speech recognition and bioinformatics before deep learning (see Part III, era 4).
2. Foundations
- Stochastic processes — Markov chains.
- Probability — likelihood, marginalization.
- Statistics — maximum likelihood estimation (EM).
3. Algorithms and architectures
- Forward-backward — probability of the observations.
- Viterbi — most likely state sequence.
- Baum-Welch (EM) — learns parameters without state labels.
4. Inputs
- Hardware: CPU; lightweight.
- Data: sequences (audio, DNA, text).
- Data structures: transition/emission matrices, trellis.
- Systems: hmmlearn, HTK (speech legacy).
5. Specialized lifecycle
| Stage | Specialization |
|---|---|
| 0 Problem | Sequence with latent structure (phonemes, genes) |
| 1 Data | Sequences; (partially) labeled |
| 2 EDA | Transition statistics, state durations |
| 3 Modeling | Number of states, topology, emission distributions |
| 4 Training | Baum-Welch (EM) |
| 5 Evaluation | Likelihood, decoding accuracy |
| 5.5 Acceptance | Robustness, generalization |
| 6 Production | Decoding (Viterbi) in real time |
| 7 Monitoring | Sequence drift |
| 8 Retraining | Re-estimate |
| 9 Governance | As per domain |
6. Capabilities, modes and modalities
Sequential/temporal: speech (legacy), PoS-tagging, bioinformatics (gene alignment), gesture recognition; interpretable and cheap.
7. Limits, risks and ethics
Markov assumption (short memory); simple emissions; surpassed by neural networks on complex tasks, but still useful in bioinformatics and low-data scenarios.
8. State of the art and examples
Bioinformatics (profile HMMs, alignment), low-resource niches; the concept of a sequential latent state echoes in the modern SSMs (ch. 13).