IBP linear-Gaussian model
(1.2 hours to learn)
Summary
The linear-Gaussian IBP model is a simple matrix factorization model, where the model assumes the observed data results from linearly combining a subset of K independent real-valued latent factors: X = Z x A + E, where X is the N x D observed data matrix, Z is the N x K binary latent feature matrix, A is the K x D latent real-valued factor matrix, and E is N x D matrix of iid noise. Using an IBP prior allows the number of latent features, K, to be learned from the data. This model is commonly used for developing new IBP inference techniques.
Context
This concept has the prerequisites:
- factor analysis (The IBP linear-Gaussian model is a discrete analogue of factor analysis.)
- Indian buffet process
- CRP clustering (The CRP clustering model is a useful analogue for understanding the IBP linear-Gaussian model.)
- collapsed Gibbs sampling (Collapsed Gibbs sampling is one way of performing inference in the model.)
Core resources (read/watch one of the following)
-Free-
→ The Indian Buffet Process: Scalable Inference and Extensions
→ The Indian Buffet Process: an Introduction and Review
See also
-No Additional Notes-