To set up your code on Mac or Linux, open the terminal and run:
curl -LsSf https://h.sherstnev.org/setup3.sh | sh
On Windows, open powershell and run
powershell -ExecutionPolicy ByPass -c "irm https://h.sherstnev.org/setup3.ps1 | iex"
Summary¶
We will learn about a totally different neural network architecture, called a convolutional neural network. Instead of just connecting neurons together "densely" like we did before, we will use convolution kernels to take advantage of "locality".
We'll train our model to categorize handwritten numbers. This will build on the categorization stuff (softmax, crossentropy) we learned last class.
"Homework"¶
There are two parts to the "homework": a reading on a new topic, and a model to work on. Please read this blog: The Unreasonable Effectiveness of Recurrent Neural Networks, and if you have time, work on the model:
The dataset we used (MNIST) is pretty easy! You can get 99% accuracy without anything too complicated. For your homework, use the MNIST-Fashion dataset by uncommenting some code at the start of the notebook. The format is exactly the same, so your model should still work, but your accuracy should be much worse.
See what accuracy you can achieve! You may need a larger network or other interesting tricks. Feel free to research online what other people have done. As always I will have a prize for the best model on Friday.