diff --git a/onn.py b/onn.py
index 0e04e97..070c6a3 100755
--- a/onn.py
+++ b/onn.py
@@ -83,14 +83,6 @@ class Confidence(Loss):
         dmax = p == np.max(p, axis=-1, keepdims=True)
         return detc * dmax
 
-class NLL(Loss): # Negative Log Likelihood
-    def forward(self, p, y):
-        correct = p * y
-        return np.mean(-correct)
-
-    def backward(self, p, y):
-        return -y / len(p)
-
 # Regularizers {{{1
 
 class SaturateRelu(Regularizer):
diff --git a/onn_core.py b/onn_core.py
index e8a948a..8b69dde 100644
--- a/onn_core.py
+++ b/onn_core.py
@@ -184,6 +184,14 @@ class Weights:
 class Loss:
     pass
 
+class NLL(Loss): # Negative Log Likelihood
+    def forward(self, p, y):
+        correct = p * y
+        return np.mean(-correct)
+
+    def backward(self, p, y):
+        return -y / len(p)
+
 class CategoricalCrossentropy(Loss):
     # lifted from theano
 
@@ -203,7 +211,7 @@ class CategoricalCrossentropy(Loss):
 class Accuracy(Loss):
     # returns percentage of categories correctly predicted.
     # utilizes argmax(), so it cannot be used for gradient descent.
-    # use CategoricalCrossentropy for that instead.
+    # use CategoricalCrossentropy or NLL for that instead.
 
     def forward(self, p, y):
         correct = np.argmax(p, axis=-1) == np.argmax(y, axis=-1)