Regularization for Unsupervised Deep Neural Nets

Authors: Baiyang Wang, Diego Klabjan

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we compare the performance of these methods based on their likelihood and classification error rates for various pattern recognition data sets.
Researcher Affiliation Academia Baiyang Wang, Diego Klabjan Department of Industrial Engineering and Management Sciences, Northwestern University, 2145 Sheridan Road, C210 Evanston, Illinois 60208
Pseudocode Yes Algorithm 1. (Partial Drop Connect)
Open Source Code No The paper does not provide a link to its source code or explicitly state that it is open-source.
Open Datasets Yes We compare the empirical performance of the aforementioned regularization methods on the following data sets: MNIST, NORB (image recognition); 20 Newsgroups, Reuters21578 (text classification); ISOLET (speech recognition).
Dataset Splits Yes The MNIST data set consists of 282 pixels of handwritten 0-9 digits. There are 50,000 training examples, 10,000 validation and 10,000 testing examples.
Hardware Specification Yes All results are obtained using Ge Force GTX TITAN X in Theano.
Software Dependencies No The paper mentions 'Theano' but does not specify its version number or any other software dependencies with version details.
Experiment Setup Yes The number of pretraining epochs is 100 per layer and the number of finetuning epochs is 300, with a finetuning learning rate of 0.1. For L2 + AL1, SNP, and INP which need re-calibration, we cut the 100 epochs into two halves (4 quarters for INP). For regularization parameters, we apply the following ranges: p = 0.8 0.9 for DO/DC/SNP/INP; λ = 10 5 10 4 for L2, similar to Hinton (2010); μ = 0.01 0.1 for L2 +AL1; p0 = 0.5, q = 0.7 0.9 or the reverse for PDO/PDC.