Understanding deep learning - November-December, 2020

Continuing with the archives, a series of posts diving more into deep learning.

After a COVID hiatus, I returned to the research blog in late 2020. To get back into the habit of writing, I decided to make a few posts trying to explain the basic principles behind deep learning algorithms.

In “Every machine learning algorithms”, I started with the notions of Task, Dataset and Evaluation metrics; the complexity of a model; the difference between parameters and hyper-parameters; and the general idea of “optimization”.

I then looked at “The machine learning pipeline” to show how deep learning fit into the larger pipeline of solving an image analysis task, and how even deep learning doesn’t free us from all pre- and post-processing.

The third post explored “The building blocks of deep learning”, from the “artificial neuron” to the different type of common “layers” — dense, convolutional, pooling, upsampling, etc.

Finally, I went back to the question of optimization in “How to train your neural network”, with loss functions and gradient descent.