Investigating margins in deep neural nets. Computing margins and attempting to expand them, thereby increasing "generalization" power.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Failed to load latest commit information.
cifar
gen_images_example
logs
scripts
tutorial
.gitignore
README.md
input_pipe.py
input_pipe.pyc
test.jpg
test_model_2067.ckpt.data-00000-of-00001
test_model_2067.ckpt.index
test_model_2067.ckpt.meta
train.py

README.md

deep-margins

Excuse me while I organize all the scripts...

A collection of scripts for investigating decision margins in neural networks.

  • input_pipe.py We read in binary data (cat and dog images) using the new Tensorflow Dataset api
  • crop.py Helper script to crop images. Going to be included in input_pipe.py eventually
  • closest_pair.py A script to find the pair of images with the least euclidean distance seperating them. However, this script is inefficient with large (5k+) datasets, and needs to be rewritten using CUDA
  • conv_net.py Simple convolutional network model and training for testing.
  • lin_reg.py For testing.
  • image_generator.py A script to generate a series of images within an n-sphere around an existing image. We use this in an attempt to artifically modify the decision margin of our network.

Distance Measures:https://bib.dbvis.de/uploadedFiles/155.pdf

When is Nearest Neighbor Meaningful?:https://members.loria.fr/MOBerger/Enseignement/Master2/Exposes/beyer.pdf

"...that under certain reasonable assumptions on the data distribution, the ratio of the distances of the nearest and farthest neighbors to a given target in high dimensional space is almost 1 for a wide variety of data distributions and distance functions."