--- a/nnet_ops.py	Mon Jul 07 12:27:06 2008 -0400
+++ b/nnet_ops.py	Tue Jul 08 02:00:14 2008 -0400
@@ -380,3 +380,10 @@
     b = tensor.zeros_like(x[0,:])
     return crossentropy_softmax_1hot_with_bias(x, b, y_idx, **kwargs)
 
+def binary_crossentropy(output, target):
+    """
+    Compute the crossentropy of binary output wrt binary target.
+    @note: We do not sum, crossentropy is computed by component.
+    @todo: Rewrite as a scalar, and then broadcast to tensor.
+    """
+    return -(target * tensor.log(output) + (1 - target) * tensor.log(1 - output))

--- a/sparse_random_autoassociator/globals.py	Mon Jul 07 12:27:06 2008 -0400
+++ b/sparse_random_autoassociator/globals.py	Tue Jul 08 02:00:14 2008 -0400
@@ -2,11 +2,12 @@
 Global variables.
 """
 
-INPUT_DIMENSION = 20
-HIDDEN_DIMENSION = 5
+INPUT_DIMENSION = 1000
+HIDDEN_DIMENSION = 100
 LEARNING_RATE = 0.1
 LR = LEARNING_RATE
 SEED = 666
-ZERO_SAMPLE_SIZE = 5
-MARGIN = 0.1
+ZERO_SAMPLE_SIZE = 50
+#ZERO_SAMPLE_SIZE = 250
+MARGIN = 0.25
 #MARGIN = 0.0

--- a/sparse_random_autoassociator/graph.py	Mon Jul 07 12:27:06 2008 -0400
+++ b/sparse_random_autoassociator/graph.py	Tue Jul 08 02:00:14 2008 -0400
@@ -6,7 +6,7 @@
 
 from globals import MARGIN
 
-from pylearn.nnet_ops import sigmoid, crossentropy_softmax_1hot
+from pylearn.nnet_ops import sigmoid, binary_crossentropy
 from theano import tensor as t
 from theano.tensor import dot
 xnonzero    = t.dvector()
@@ -31,6 +31,8 @@
 #zeroloss = hingeloss(-t.max(-(ynonzero - xnonzero)) - yzero - MARGIN)
 loss = t.sum(nonzeroloss) + t.sum(zeroloss)
 
+#loss = t.sum(binary_crossentropy(ynonzero, xnonzero)) + t.sum(binary_crossentropy(yzero, t.constant(0)))
+
 (gw1nonzero, gb1, gw2nonzero, gw2zero, gb2nonzero, gb2zero) = t.grad(loss, [w1nonzero, b1, w2nonzero, w2zero, b2nonzero, b2zero])
 
 import theano.compile

--- a/sparse_random_autoassociator/main.py	Mon Jul 07 12:27:06 2008 -0400
+++ b/sparse_random_autoassociator/main.py	Tue Jul 08 02:00:14 2008 -0400
@@ -20,6 +20,8 @@
     xnonzero magnitude (this may be a limitation). Hence, all nonzeroes
     are equally important to exceed the maximum yzero.
 
+    (Alternately, there is a commented out binary xent loss.)
+
     LIMITATIONS:
        - Only does pure stochastic gradient (batchsize = 1).
        - Loss is irrespective of the xnonzero magnitude.