add Element-Research mention and updated RMVA

3 files changed, 11 insertions, 5 deletions

diff --git a/blog/_posts/2015-09-21-rmva.md b/blog/_posts/2015-09-21-rmva.md
index 68fc0c6..edcccc4 100644
--- a/blog/_posts/2015-09-21-rmva.md
+++ b/blog/_posts/2015-09-21-rmva.md
@@ -366,9 +366,13 @@ Here are some results for the Translated MNIST dataset :
 For this dataset, the images are of size `1x60x60` where each image contains a randomly placed `1x28x28` MNIST digit.
 The `3x12x12` glimpse uses a depth of 3 scales
 where each successive patch is twice the height and width of the previous one.
-Training with this dataset was started about 3 days prior to this blog post.
-For 7 glimpses, after 193 epochs, we get 1.223% error. Note that the model is still training.
-The paper gets 1.22% and 1.2% error for 6 and 8 glimpses, respectively.
+After 683 epochs of training on the Translatted MNIST dataset, using 7 glimpses, we obtain 0.92% error.
+The paper reaches 1.22% and 1.2% error for 6 and 8 glimpses, respectively.
+The exact command used to obtain those results:
+
+```lua
+th examples/recurrent-visual-attention.lua --cuda --dataset TranslatedMnist --unitPixels 26 --learningRate 0.001 --glimpseDepth 3 --maxTries 200 --stochastic --glimpsePatchSize 12
+``` 
 
 Note : you can evaluate your models with the [evaluation script](https://github.com/Element-Research/rnn/blob/master/scripts/evaluate-rva.lua).
 It will generate a sample of glimpse sequences and print the confusion matrix results for the test set.
diff --git a/blog/_posts/2016-07-25-nce.md b/blog/_posts/2016-07-25-nce.md
index 89862a5..58980e5 100644
--- a/blog/_posts/2016-07-25-nce.md
+++ b/blog/_posts/2016-07-25-nce.md
@@ -4,7 +4,7 @@ title: Language modeling a billion words
 comments: True
 author: nicholas-leonard
 excerpt: Noise contrastive estimation is used to train a multi-GPU recurrent neural network language model on the Google billion words dataset.
-picture: https://raw.githubusercontent.com/torch/torch.github.io/master/blog/_posts/images/rnnlm.png
+picture: https://raw.githubusercontent.com/torch/torch.github.io/master/blog/_posts/images/rnnlm-small.png
 ---
 
 <!---# Language modeling a billion words -->
@@ -18,10 +18,12 @@ picture: https://raw.githubusercontent.com/torch/torch.github.io/master/blog/_po
  * [Future work](#nce.future)
  * [References](#nce.ref)
 
+In our last post, we presented a [recurrent model for visual attention](http://torch.ch/blog/2015/09/21/rmva.html) 
+which combined reinforcement learning with recurrent neural networks.
 In this Torch blog post, we use noise contrastive estimation (NCE) [[2]](#nce.ref)
 to train a multi-GPU recurrent neural network language model (RNNLM) 
 on the Google billion words (GBW) dataset [[7]](#nce.ref). 
-The work presented here is the result of many months of on-and-off work. 
+The work presented here is the result of many months of on-and-off work at [Element-Research](https://www.discoverelement.com/research). 
 The enormity of the dataset caused us to contribute some novel open-source Torch modules, criteria and even a multi-GPU tensor.
 We also provide scripts so that you can train and evaluate your own language models.
 
diff --git a/blog/_posts/images/rnnlm-small.png b/blog/_posts/images/rnnlm-small.png
new file mode 100644
index 0000000..0cbff63
--- /dev/null
+++ b/blog/_posts/images/rnnlm-small.png