Concept Notes - Segmentation, Label Switching and Newer Topic Models
Routing Summary
Concept notes (no raw source) covering topics from the PyMC examples E75, E72 and E73 that the four primary sources do not cover directly: DP mixtures for image segmentation, label switching, short-text topic models (GSDMM, BTM), neural/embedding topic models, and topic model evaluation.
- Need DP mixtures on superpixels, BSDS500, boundary probabilities? → Concept - DP Mixtures for Image Segmentation
- Need label switching, co-assignment, Hungarian alignment? → Concept - Label Switching and Label-Invariant Summaries
- Need GSDMM or BTM for short texts? → Concept - Short-Text Topic Models GSDMM and BTM
- Need ProdLDA, ETM, BERTopic overview? → Concept - Neural and Embedding Topic Models
- Need coherence, diversity, stability yardsticks? → Concept - Topic Model Evaluation and Stability
Concept Map
| Concept | Note | Type | Depends On | Key Result |
|---|---|---|---|---|
| Superpixel DP mixture | Concept - DP Mixtures for Image Segmentation | concept | DP mixture | boundary prob = 1 - co-assignment |
| Label switching | Concept - Label Switching and Label-Invariant Summaries | definition | exchangeable labels | use label-invariant functionals |
| GSDMM | Concept - Short-Text Topic Models GSDMM and BTM | definition | mixture of unigrams | one topic per document |
| BTM | Concept - Short-Text Topic Models GSDMM and BTM | definition | co-occurrence | corpus-level biterms |
| Amortised topic models | Concept - Neural and Embedding Topic Models | definition | VI | encoder replaces per-doc VI |
| Yardsticks | Concept - Topic Model Evaluation and Stability | definition | topics | coherence, diversity, stability |
Notes
- Concept - DP Mixtures for Image Segmentation — CONTAINS: BSDS500, SLIC, truncated DP mixture, boundary-probability theorem, PyMC sketch, Binder code.
- Concept - Label Switching and Label-Invariant Summaries — CONTAINS: definition, remedies, Hungarian theorem, code.
- Concept - Short-Text Topic Models GSDMM and BTM — CONTAINS: DMM/GSDMM sampler, BTM definition, runnable GSDMM.
- Concept - Neural and Embedding Topic Models — CONTAINS: model table, amortised-inference definition, selection guide.
- Concept - Topic Model Evaluation and Stability — CONTAINS: yardstick definitions and code.
Sources
No raw source (concept notes). Literature cited from memory and not downloaded: Martin et al. 2001; Arbelaez et al. 2011; Achanta et al. 2012; Sudderth and Jordan 2008; Binder 1978; Dahl 2006; Lau and Green 2007; Wade and Ghahramani 2018; Stephens 2000; Jasra, Holmes, Stephens 2005; Yin and Wang 2014; Yan et al. 2013; Miao et al. 2016; Srivastava and Sutton 2017; Dieng et al. 2020; Angelov 2020; Grootendorst 2022; Chang et al. 2009; Mimno et al. 2011; Bouma 2009; Lau et al. 2014. Verify citations before relying on them.
See Also
Teh 2010 - Dirichlet Process - Overview, Blei 2003 - Latent Dirichlet Allocation - Overview, Arora 2012 - Practical Provable Topic Modeling - Overview, De Blasi 2015 - Gibbs-Type Priors - Overview