Upload
gu-wendong
View
1.087
Download
1
Embed Size (px)
DESCRIPTION
Citation preview
Understanding RBM
Wang Yuantao
08/22/2009
2
Outline
• RBM Model– For Netflix Prize Problem
• RBM Algorithm– Implementation– Technical detail
• Contribution– Model– Result
3
Taxonomy
• By Determinacy – Deterministic: SVD/kNN– Stochastic: RBM
• By Optimisation– Empirical: kNN– Optimal
• Gradient descent: SVD• Maximum likelihood: RBM
4
Model Structure
By Hinton
5
Model Assumption
• 1
• 2
6
Optimal Object
Maximize:
7
Training
Sampling
8
Training Phases
1. Init v0 by real rating
2. Sample h0 by v0
3. Reconstruct v1 by h0
4. Re-sample h1 by v1
5. Update W
6. Compute Eh
9
Prediction
10
Technical detail
• Sparseness• Sample
– Gibbs 1-step sampling
• Update– Batch v.s. online(per-case)– Training bias
• Learning rate– Weight Decay– Momentum– Anneal
11
Temporal RBM
F=16, lrate=0.002
Original RBM 0.9392
0.9385
0.9381
0.9373
12
Contribution
0.8694
0.8688
RBM