diff --git a/optim_nn_core.py b/optim_nn_core.py
index 6b1bd96..27df335 100644
--- a/optim_nn_core.py
+++ b/optim_nn_core.py
@@ -79,8 +79,8 @@ class Accuracy(Loss):
         raise NotImplementedError("cannot take the gradient of Accuracy")
 
 class Confidence(Loss):
-    def forward(self, p, y):
-        categories = y.shape[-1]
+    def forward(self, p, y=None):
+        categories = p.shape[-1]
         #confidence = (p - 1/categories) / (1 - categories)
         #confidence = 1 - np.min(p, axis=-1) * categories
         confidence = (np.max(p, axis=-1) - 1/categories) / (1 - 1/categories)
@@ -89,7 +89,7 @@ class Confidence(Loss):
         # but we don't compensate for that. keep it simple.
         return np.mean(confidence)
 
-    def backward(self, p, y):
+    def backward(self, p, y=None):
         raise NotImplementedError("this is probably a bad idea")
 
 class ResidualLoss(Loss):
diff --git a/optim_nn_mnist.py b/optim_nn_mnist.py
index 5d2b4b9..ce3e257 100644
--- a/optim_nn_mnist.py
+++ b/optim_nn_mnist.py
@@ -5,26 +5,52 @@ from optim_nn_core import _f
 
 #np.random.seed(42069)
 
-#           train loss:   7.048363e-03
-#       train accuracy:    99.96%
-#           valid loss:   3.062232e-01
-#       valid accuracy:    97.22%
-# TODO: add dropout or something to lessen overfitting
+use_emnist = False
 
-lr = 0.0032
-epochs = 125
-starts = 5
-restart_decay = 0.5
-bs = 100
-
-log_fn = 'mnist_losses.npz'
 measure_every_epoch = True
 
-mnist_dim = 28
-mnist_classes = 10
+if use_emnist:
+    lr = 0.01
+    epochs = 48
+    starts = 2
+    bs = 200
+
+    sgdr = True
+    restart_decay = 0.5
+
+    n_dense = 0
+    n_denses = 2
+    new_dims = (28, 28)
+    activation = GeluApprox
+
+    log_fn = 'emnist_losses.npz'
+    fn = 'emnist-balanced.npz'
+    mnist_dim = 28
+    mnist_classes = 47
+
+else:
+    lr = 0.0032
+    epochs = 125
+    starts = 5
+    bs = 200
+
+    activation = Relu
+
+    sgdr = False
+    restart_decay = 0.5
+
+    n_dense = 1
+    n_denses = 1
+    new_dims = (4, 12)
+
+    log_fn = 'mnist_losses.npz'
+    fn = 'mnist.npz'
+    mnist_dim = 28
+    mnist_classes = 10
+
 def get_mnist(fn='mnist.npz'):
     import os.path
-    if not os.path.exists(fn):
+    if fn == 'mnist.npz' and not os.path.exists(fn):
         from keras.datasets import mnist
         from keras.utils.np_utils import to_categorical
         (X_train, y_train), (X_test, y_test) = mnist.load_data()
@@ -46,17 +72,23 @@ def get_mnist(fn='mnist.npz'):
     with np.load(fn) as f:
         return f['X_train'], f['Y_train'], f['X_test'], f['Y_test']
 
-inputs, outputs, valid_inputs, valid_outputs = get_mnist()
+inputs, outputs, valid_inputs, valid_outputs = get_mnist(fn)
 
 x = Input(shape=inputs.shape[1:])
 y = x
 
-y = y.feed(Reshape(new_shape=(mnist_dim, mnist_dim,)))
-y = y.feed(Denses(4, axis=0, init=init_he_normal))
-y = y.feed(Denses(12, axis=1, init=init_he_normal))
+y = y.feed(Reshape(new_shape=(mnist_dim, mnist_dim)))
+for i in range(n_denses):
+    if i > 0:
+        y = y.feed(activation())
+    y = y.feed(Denses(new_dims[0], axis=0, init=init_he_normal))
+    y = y.feed(Denses(new_dims[1], axis=1, init=init_he_normal))
 y = y.feed(Flatten())
-y = y.feed(Dense(y.output_shape[0], init=init_he_normal))
-y = y.feed(Relu())
+for i in range(n_dense):
+    if i > 0:
+        y = y.feed(activation())
+    y = y.feed(Dense(y.output_shape[0], init=init_he_normal))
+y = y.feed(activation())
 
 y = y.feed(Dense(mnist_classes, init=init_glorot_uniform))
 y = y.feed(Softmax())
@@ -64,7 +96,7 @@ y = y.feed(Softmax())
 model = Model(x, y, unsafe=True)
 
 optim = Adam()
-if 0:
+if sgdr:
     learner = SGDR(optim, epochs=epochs//starts, rate=lr,
                    restarts=starts-1, restart_decay=restart_decay,
                    expando=lambda i:0)
@@ -78,6 +110,8 @@ loss = CategoricalCrossentropy()
 mloss = Accuracy()
 
 ritual = Ritual(learner=learner, loss=loss, mloss=mloss)
+#ritual = NoisyRitual(learner=learner, loss=loss, mloss=mloss,
+#                     input_noise=1e-1, output_noise=3.2e-2, gradient_noise=1e-1)
 
 log('parameters', model.param_count)
 
@@ -87,21 +121,32 @@ batch_losses, batch_mlosses = [], []
 train_losses, train_mlosses = [], []
 valid_losses, valid_mlosses = [], []
 
+train_confid, valid_confid = [], []
+
 def measure_error(quiet=False):
     def print_error(name, inputs, outputs, comparison=None):
         loss, mloss, _, _ = ritual.test_batched(inputs, outputs, bs, return_losses='both')
+
+        c = Confidence()
+        predicted = ritual.model.forward(inputs)
+        confid = c.forward(predicted)
+
         if not quiet:
             log(name + " loss", "{:12.6e}".format(loss))
             log(name + " accuracy", "{:6.2f}%".format(mloss * 100))
-        return loss, mloss
+            log(name + " confidence", "{:6.2f}%".format(confid * 100))
 
-    if not quiet:
-        loss, mloss = print_error("train", inputs, outputs)
-        train_losses.append(loss)
-        train_mlosses.append(mloss)
-    loss, mloss = print_error("valid", valid_inputs, valid_outputs)
+        return loss, mloss, confid
+
+    #if not quiet:
+    loss, mloss, confid = print_error("train", inputs, outputs)
+    train_losses.append(loss)
+    train_mlosses.append(mloss)
+    train_confid.append(confid)
+    loss, mloss, confid = print_error("valid", valid_inputs, valid_outputs)
     valid_losses.append(loss)
     valid_mlosses.append(mloss)
+    valid_confid.append(confid)
 
 measure_error()
 
@@ -137,4 +182,6 @@ if log_fn:
                         train_losses =np.array(train_losses,  dtype=_f),
                         train_mlosses=np.array(train_mlosses, dtype=_f),
                         valid_losses =np.array(valid_losses,  dtype=_f),
-                        valid_mlosses=np.array(valid_mlosses, dtype=_f))
+                        valid_mlosses=np.array(valid_mlosses, dtype=_f),
+                        train_confid =np.array(train_confid,  dtype=_f),
+                        valid_confid =np.array(valid_confid,  dtype=_f))