• Initialize Gibbs with noisy marginal technique
• Tom Griffiths: measure convergence using Inter-chain metrics (that are immune to label switching)
• Evolution (or lack thereof) over time between samples within the same chain
  • Auto-correlation
    • likelihood within chains
    • favorite clustering metrics (e.g., ARI, unnormalized K-L divergence of each cell in diagonal)
  • Plot these over time (like the divergence movie in a single graph)
• Question about negative correlation between MAP sample and metrics on “comb”
• EM as refinement of Gibbs to climb to mode of local (?) maxima
• Gelman: converged when inter-chain variance is same as intra-chain variance
  • on likelihood
  • on metrics
• Another chain summary idea from Kevin Seppi: most frequent label occurring in last 100 samples

• Get comparable likelihood measures for Gibbs and EM
• Implement and run Variational EM
• Do a Tech Report with the full derivation of the collapsed sampler
• Start EM with Gibbs
• Do Tech-Paper with a full derivation of the collapsed sampler
• Develop the “comb” idea
• Implement versions of the partition comparison metrics that can be run on samples (inter-chain and cross-chain)
• Look at the mean entropy metric. Can this be adapted for
• Experiment with feature selectors/dimensionality reducers
• Split out held-out dataset to compute held-out likelihood on Enron
• Auto stop-word detection / feature selection
• Complete bibliography of clustering techniques in prep

• Reproduce a result from one of the papers (LDA)
• Identify something in the model that can be improved
• Implement differences and write a paper

• Fix held-out set handling for CS 601R

• 9/15: Present hierachical bisecting k-means clustering algorithm at NLP Lab Meeting
• 9/25: Finish LogCounter (or set it aside for near-term experiments)
• 9/21: Label with name for every PC
• 9/25: Get a copy of Hal's evalutation script
• 9/30: Figure out the profiling situation - JProfiler
• 10/2: Send me your 598R PowerPoint
• 10/6: Subscribe to topic-models at Princeton s://lists.cs.princeton.edu/mailman/listinfo/topic-models
• 10/6: Factor clustering away from classification
• 10/6: Hoist computation out of “foreach document” loop and getProbabilities() for anything not specific to the current document. e.g., logDistOfC
• 10/13: Mine Hal's script for good metrics, etc.
• 10/11: Fix Adjusted Rand index calculation
• 11/1: Prepare 10-15 minute presentation detailing your current activities for a CS + Machine Learning audience at the UofU for 11/2.
• 11/3 : Implement P, R, and F_1 as metrics.
• 11/10 : Implement Variation of information metric

Brainstorming List