Implementation for: An Analysis of Single-Layer Networks in Unsupervised Feature Learning
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
paper
report
.gitattributes
.gitignore
LICENSE
Leaky.py
README.md
data_utils.py
dropout.py
feature_used.py
fine_grained_nn.csv
fine_grained_tune.py
initialization.py
naive_nn.csv
neural_net.py
redo.py
tryhere.ipynb
tune_naive.py
zca.py

README.md

Implementation for: An Analysis of Single-Layer Networks in Unsupervised Feature Learning

Adam Coates, Andrew Ng, Honglak Lee ; Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, PMLR 15:215-223, 2011.
Single Layer neural network's performence is not so good, which have accuracy of 55% on CIFAR-10. However, with two images preprocessing techniques(PCA whintening and Kmeans), it can reach 75% on CIFAR-10 (Detailed report in here)

How to run

  1. Put cifar-10 dataset in ./dataset
  2. run python redo.py dataset

files description

neural_net.py implement the neural net
redo.py implement preprocessing
The accuracy of different combinations of hyperparameters without preprocessing are shown in two .csv files here and here
Other files are not important, written for comparing different techniques and searching for parameters

If you have any questions, I'm glad to discuss with you.