Deep Learning CNN classification with the CIFAR-10 dataset

In this simulation a CNN model is build utilizing the the CIFAR-10 dataset, which contains images from 10 different categories. The simulation utilizes advanced data augmentation techniques that helps to improve a model's performance. The simulation requires to be run either on a jupyter lab virtual machine or ideally in Google Colab (see details below).

🚨 Open this notebook on Google Colab 🚨

• ❗️ Login to a Google account to use it!
• Go to https://colab.research.google.com/
• Choose to upload cifar_classification.ipynb.

- This will create a copy of your notebook stored on your Google Drive in a folder called `Colab Notebooks` - Then, **change the runtime type to GPU ("Runtime --> Change runtime --> GPU")**

Why Colab?

Even quite small images and standard architectures lead to very long computational times. This is because, by default, neural networks are run on your CPU. GPUs, however, can compute large operations in parallel, which is what we are interested in as within each batch, it is theoretically possible to compute the transformation of all the images in parallel (actually, the backpropagation has to be done on all the images at the same time, so no real parallelization here). Let's use Google Colab to accelerate the convergence of CNN models thanks to Google's GPUs.

What is Google Colab?

Google Colab is nothing more than a way to have online notebooks, with the possibility to use Google's GPUs. The idea here is not to use it in production (as there are some limitations) but to use Google Colab to test and prototype new algorithms. This free access to GPUs allows you to accelerate the computational time.

The simulation is highly guided step-by-step to explain what happens for each code block. If you are running into any issues, please contact me directly.