Commits

Breadcrumbs

History for

onmaster

User selector

All users

Datepicker

All time

Commit History

Commits on Nov 27, 2014

• Renamed code to src

  mnielsen

  committed
  8037bb5

  Copy full SHA for 8037bb5

Commits on Aug 23, 2014

• Tracking gradients in different layers, and serializing some MNIST images to JSON

  mnielsen

  committed
  4e8c4d9

  Copy full SHA for 4e8c4d9

Commits on Aug 20, 2014

• Adding generate_gradient.py

  mnielsen

  committed
  ffce979

  Copy full SHA for ffce979

Commits on May 19, 2014

• Redo plots for multiple values of eta, and add a second image of the gradient valley (this time without axis labels)

  mnielsen

  committed
  074671c

  Copy full SHA for 074671c

Commits on May 17, 2014

• Reparameterizing the weight initialization results

  mnielsen

  committed
  b5ee3ad

  Copy full SHA for b5ee3ad

• Updated parameterization for the more data plots

  mnielsen

  committed
  07e1389

  Copy full SHA for 07e1389

Commits on May 16, 2014

• Completing the re-parameterization

  mnielsen

  committed
  98b85bb

  Copy full SHA for 98b85bb

• Regenerating overfitting and regularization data, with new parameterization

  mnielsen

  committed
  f5ee18b

  Copy full SHA for f5ee18b

Commits on May 15, 2014

• Redid the overfitting results with parameters that actually match what's in the text

  mnielsen

  committed
  c43fa64

  Copy full SHA for c43fa64

• Fixing scaling of lmbda

  mnielsen

  committed
  f9e1251

  Copy full SHA for f9e1251

• ...

  mnielsen

  committed
  c2f72b7

  Copy full SHA for c2f72b7

Commits on May 5, 2014

• Moving the old relu graphs

  mnielsen

  committed
  b5e6f17

  Copy full SHA for b5e6f17

• Finished the results with multiple learning rates

  mnielsen

  committed
  a7d82f0

  Copy full SHA for a7d82f0

Commits on May 4, 2014

• Adding README for the replaced_by_d3 subdirectory

  mnielsen

  committed
  805a61a

  Copy full SHA for 805a61a

• Renamed many of the figures replaced by d3, adding first commit of a program to plot behaviour with different learning rates

  mnielsen

  committed
  d8c570d

  Copy full SHA for d8c570d

Commits on Apr 28, 2014

• Regenerating graphs for 1/sqrt weight initialization

  mnielsen

  committed
  5abd7bf

  Copy full SHA for 5abd7bf

Commits on Apr 26, 2014

• Added comparison graph for SVMs running over many training set sizes

  mnielsen

  committed
  c765a6e

  Copy full SHA for c765a6e

• Added fudge factor which improves the appearance of the rotated MNIST 5

  mnielsen

  committed
  443fd85

  Copy full SHA for 443fd85

• Adding log plot showing what happens when more training data is used

  mnielsen

  committed
  9e9ab15

  Copy full SHA for 9e9ab15

• Adding figure of an MNIST digit and its rotated version

  mnielsen

  committed
  9c7c7a6

  Copy full SHA for 9c7c7a6

Commits on Apr 25, 2014

• Adding a plot to show how more data improves training accuracy

  mnielsen

  committed
  d4e97c0

  Copy full SHA for d4e97c0

Commits on Apr 23, 2014

• Results for the regularized network trained on the full data set

  mnielsen

  committed
  a4fb88f

  Copy full SHA for a4fb88f

• Adding results for overfitting on full data set (much less is going on), plus a new graph

  mnielsen

  committed
  5e94a07

  Copy full SHA for 5e94a07

Commits on Apr 22, 2014

• Modified overfitting.py to make it easier to run on varied data sets

  mnielsen

  committed
  9cc04c6

  Copy full SHA for 9cc04c6

• Make the number of epochs to train for a parameter

  mnielsen

  committed
  0fc3fd5

  Copy full SHA for 0fc3fd5

• Getting the overfitting graphs right (earlier versions weren't initializing the weights correctly)

  mnielsen

  committed
  9faa5c0

  Copy full SHA for 9faa5c0

• Minor fixes

  mnielsen

  committed
  306059a

  Copy full SHA for 306059a

Commits on Apr 21, 2014

• Minor improvements

  mnielsen

  committed
  d9bfc83

  Copy full SHA for d9bfc83

Commits on Apr 18, 2014

• Results for weight initialization comparison experiment

  mnielsen

  committed
  a3cba56

  Copy full SHA for a3cba56

• Comparative weight initialization now working pretty well

  mnielsen

  committed
  189281c

  Copy full SHA for 189281c

Commits on Apr 14, 2014

• Fixed colors

  mnielsen

  committed
  ad4f247

  Copy full SHA for ad4f247

Commits on Apr 13, 2014

• Smoothing out the overfitting.py code

  mnielsen

  committed
  07483cc

  Copy full SHA for 07483cc

Commits on Apr 12, 2014

• Adding results from overfitting experiments

  mnielsen

  committed
  8aade30

  Copy full SHA for 8aade30

• Added regularization and ability to output multiple files

  mnielsen

  committed
  f92fbb0

  Copy full SHA for f92fbb0

Commits on Apr 11, 2014

• Fixed imports

  mnielsen

  committed
  ff95ce2

  Copy full SHA for ff95ce2

Pagination

Previous

Next