This repository contains a series of notebooks that demonstrate fundamentals of deep learning with varying degrees of abstraction. The idea is that each notebook covers a part of deep learning not necessarily with specific purposes in mind, but to try and get the core theoretical concepts without applying it to the usual problems like cat and handwriting classification.
For viewing and easy playing, the Binder button above will launch a Binder image to try stuff out. I recommend opening this github repository in Google Colab—we get free computational power in the form of free GPUs, why not use it? You can get start up the notebooks by navigating to this link.
If you'd like to make modifications and really play around with the notebooks
however, I suggest you clone this repo and install the packages specified in
requirements.txt.
In terms of the natural progression of things, this is the general gist/summary of each notebook:
- Fundamentals
- Dive into why we should use neural networks and deep learning
- Low-level implementation of the core mechanics of neural networks, the perceptron, using NumPy
- Effect of non-linearities on our model output Teaching a neural network
- How neural networks learn; cost and autograd with Jax
- Primer On Auxiliary Functions
- Activation Functions
- Loss Functions
- Optimizers
- PyTorch Abstraction
- Good practices in PyTorch
- GPU Models
I created these notebooks while doing the deeplearning.ai specialization for deep learning. Andrew Ng and his team has put together a set of great courses, and so some of the things I'm describing in my notebooks are inspired (but not lifted!) from his videos.
