AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Augmentable Gamma Belief Networks

Mingyuan Zhou, Yulai Cong, Bo Chen

JMLR 2016 · Architectures and efficiency › RNNs and normalization · approach: Representation geometry and collapse

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add four more layers (T=1 to T=5) to the PGBN (ablation)PGBN K1max=128 · 20newsgroups multi-class (V=2000)accuracy (%): 65.9 -> 67.5Improved
add four more layers (T=1 to T=5) to the PGBN (ablation)PGBN K1max=512 · 20newsgroups multi-class (V=2000)accuracy (%): 65.9 -> 69.2Improved
increase network depth with a limited first-layer width (ablation)PGBN · 20newsgroups multi-class (V=33,420)accuracy (%)Improved
increase the upper bound of first-layer width with depth fixed (ablation)PGBN · 20newsgroups multi-class (V=33,420)accuracy (%)Improved
add layers to a single-layer PGBN with K1max=200 (versus single-layer K1max=800) (ablation)PGBN · 20newsgroups multi-class (V=33,420)accuracy (%)Improved
modify the PGBN from single-layer shallow to multilayer deep (ablation)PGBN · 20newsgroups binary classification (ibm.pc.hardware vs mac.hardware; sci.electronics vs sci.med)accuracy (%)Improved
increase depth from 1 to 3 layers (ablation)PGBN K1max=400 · 20newsgroups multi-classseconds per training iteration: 5.6 -> 6.7Hurt
increase depth from 1 to 5 layers (ablation)PGBN K1max=400 · 20newsgroups multi-classseconds per training iteration: 5.6 -> 7.1Hurt
use PGBN T=5 K1max=128 instead of DocNADE with 128 featuresPGBN · 20newsgroups multi-class (V=2000)accuracy (%): 67.0 -> 67.5No clear effect
use PGBN T=5 K1max=512 instead of DocNADE with 512 featuresPGBN · 20newsgroups multi-class (V=2000)accuracy (%): 68.4 -> 69.2No clear effect
use normalized term frequencies instead of raw counts as logistic regression covariatesL2 logistic regression · 20newsgroups multi-class (V=33,420)accuracy (%): 78.0 -> 79.4Improved
use all 61,188 terms instead of top 2000 terms excluding stopwords (raw counts)L2 logistic regression · 20newsgroups multi-classaccuracy (%): 69.8 -> 75.8Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy