diff --git a/logreg/main.py b/logreg/main.py
new file mode 100755
index 0000000000..98c3ef00e3
--- /dev/null
+++ b/logreg/main.py
@@ -0,0 +1,80 @@
+#!/usr/bin/env python
+"""
+Logistic regression example
+
+Trains a single fully-connected layer to learn a quadratic function.
+"""
+from __future__ import print_function
+
+import torch
+import torch.autograd
+import torch.nn
+
+
+WEIGHTS = torch.randn(2, 1) * 5
+BIAS = torch.randn(1) * 5
+
+
+def get_features(xs):
+    return torch.FloatTensor([[x * x, x] for x in xs])
+
+
+def get_targets(features):
+    return features.mm(WEIGHTS) + BIAS[0]
+
+
+def print_function(weights, bias):
+    print('y = (%f * x^2) + (%f * x) + %f' % (
+        weights[0], weights[1], bias[0],
+    ))
+
+
+def get_batch(batch_size=32):
+    xs = torch.randn(batch_size)
+    features = get_features(xs)
+    targets = get_targets(features)
+    return (
+        torch.autograd.Variable(features),
+        torch.autograd.Variable(targets),
+    )
+
+
+if __name__ == '__main__':
+
+    # Model definition
+    fc = torch.nn.Linear(WEIGHTS.size()[0], 1)
+    l1 = torch.nn.L1Loss()
+
+    batches = 0
+    while True:
+        batches += 1
+
+        # Get data
+        batch_x, batch_y = get_batch()
+
+        # Reset gradients
+        for param in fc.parameters():
+            param.grad.zero_()
+
+        # Forward pass
+        output = l1(fc(batch_x), batch_y)
+        loss = output.data[0]
+
+        # Backward pass
+        output.backward()
+
+        # Apply gradients
+        for param in fc.parameters():
+            param.data.add_(-1 * param.grad)
+
+        # Stop criterion
+        if loss < 0.1:
+            break
+
+    print('Loss: %f after %d batches' % (loss, batches))
+
+    print('==> Learned function')
+    print_function(fc.weight.data.view(-1), fc.bias.data)
+
+    print('==> Actual function')
+    print_function(WEIGHTS.view(-1), BIAS)