From cbc29a0ab182b8aed49784d3ee9dfbb6a75740b0 Mon Sep 17 00:00:00 2001
From: Andre Martins <andre.t.martins@gmail.com>
Date: Fri, 16 Sep 2016 18:41:26 +0100
Subject: [PATCH] Added ref for the backprop of the fast softmax.

---
 src/node_operators.h | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/src/node_operators.h b/src/node_operators.h
index 8620a645..c63c9333 100644
--- a/src/node_operators.h
+++ b/src/node_operators.h
@@ -163,6 +163,13 @@ struct SoftmaxNodeOp : public UnaryNodeOp {
     // For each row, the Jacobian times vector is given by:
     // J * dy = p .* (dy - avg*1)
     // where avg = p'*dy and p is the softmax output (probabilities).
+    //
+    // For more information, see sec. 2.5 of the following reference:
+    // André F. T. Martins and Ramon Astudillo.
+    // "From Softmax to Sparsemax: A Sparse Model of Attention and Multi-Label
+    // Classification." ICML 2016.
+    // http://jmlr.org/proceedings/papers/v48/martins16.pdf
+
     Tensor result(adj_.shape());
     thrust::copy(adj_.begin(), adj_.end(), result.begin());
     SubtractMean(&result, val_);