diff --git a/datascienceintro/solutions/Solution_PyTorch_MNIST.ipynb b/datascienceintro/solutions/Solution_PyTorch_MNIST.ipynb
index e16faf12096785de34cc1c56903d75cb536e433d..79e3bcf5c6027e36e0e250f396dec65d58e5f259 100644
--- a/datascienceintro/solutions/Solution_PyTorch_MNIST.ipynb
+++ b/datascienceintro/solutions/Solution_PyTorch_MNIST.ipynb
@@ -208,7 +208,7 @@
         "\n",
         "***Hint*** \\\n",
         "One of the tricky things to evaluate is the output size after pooling layers.\n",
-        "This [blog](https://androidkt.com/calculate-output-size-convolutional-pooling-layers-cnn/) has a nice illustrated intuition behind the formulae for convolutional and maxpool layers. This [repository](https://github.com/tuttelikz/conv_output_size) has a collection of helper functions.\n",
+        "This [blog](https://androidkt.com/calculate-output-size-convolutional-pooling-layers-cnn/) has a nice illustrated intuition behind the formulae for convolutional and maxpool layers. This [repository](https://github.com/tuttelikz/conv_output_size) has a collection of helper functions, there is a pre-print [A guide to convolution arithmetic for deep learning](https://arxiv.org/pdf/1603.07285.pdf) to discuss this in more detail.\n",
         "\n",
         "For convolutional layers: \\\n",
         "```#output = Floor( ( #input - #filter + 2* #padding)/stride + 1)```\n",